Technological University Dublin

ARROW@TU Dublin

Articles School of Electrical and Electronic Engineering

2020

Applications of Artificial Intelligence to Cryptography

Jonathan Blackledge
Technological University Dublin, [email protected]

Napo Mosola
University of KwaZulu-Natal

Follow this and additional works at: https://arrow.tudublin.ie/engscheleart2

Part of the Computer Engineering Commons, and the Electrical and Computer Engineering Commons

Recommended Citation
Blackledge, J. & Mosola, N. (2020) Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning & Artifical Intellengence6th June 2020. doi:10.14738/tmlai.83.8219

This Article is brought to you for free and open access by the School of Electrical and Electronic Engineering at
ARROW@TU Dublin. It has been accepted for inclusion in Articles by an authorized administrator of ARROW@TU
Dublin. For more information, please contact [email protected], [email protected],
[email protected].
 Volume 8 No 3

Applications of Artificial Intelligence to Cryptography

Jonathan Blackledge and 6,7Napo Mosola
1,2,3,4,5
1
Stokes Professor, Science Foundation Ireland;
2
Honorary Professor, School of Electrical and Electronic Engineering, Technological University Dublin;
3
Visiting Professor, Faculty of Arts, Science and Technology, University of Wales, Wrexham Glyndwr;
4
Professor Extraordinaire, Department of Computer Science, University of Western Cape;
5
Honorary Professor and 6Research Associate, School of Mathematics, Statistics and Computer Science,
University of KwaZulu-Natal; 7Lecturer in Computer Science, National University of Lesotho.
jonathan.blackledge@TUDublin; [email protected]

ABSTRACT

This paper considers some recent advances in the field of Cryptography using Artificial Intelligence (AI). It
specifically considers the applications of Machine Learning (ML) and Evolutionary Computing (EC) to
analyze and encrypt data. A short overview is given on Artificial Neural Networks (ANNs) and the principles
of Deep Learning using Deep ANNs. In this context, the paper considers: (i) the implementation of EC and
ANNs for generating unique and unclonable ciphers; (ii) ML strategies for detecting the genuine
randomness (or otherwise) of finite binary strings for applications in Cryptanalysis. The aim of the paper
is to provide an overview on how AI can be applied for encrypting data and undertaking cryptanalysis of
such data and other data types in order to assess the cryptographic strength of an encryption algorithm,
e.g. to detect patterns of intercepted data streams that are signatures of encrypted data. This includes
some of the authors’ prior contributions to the field which is referenced throughout. Applications are
presented which include the authentication of high-value documents such as bank notes with a
smartphone. This involves using the antenna of a smartphone to read (in the near field) a flexible radio
frequency tag that couples to an integrated circuit with a non-programmable coprocessor. The
coprocessor retains ultra-strong encrypted information generated using EC that can be decrypted on-line,
thereby validating the authenticity of the document through the Internet of Things with a smartphone.
The application of optical authentication methods using a smartphone and optical ciphers is also briefly
explored.
Keywords: Artificial Intelligence, Artificial Neural Networks, Evolutionary Computing, Machine Learning,
Cryptography, Cryptanalysis, Radio Frequency Identification, Optical Authentication.

1 Introduction
The term Artificial Intelligence (AI) covers a range of methodologies and applications that are designed to
enable a computer to undertake tasks that are conventionally the domain of human intelligence. AI is
‘the ability of a digital computer or a computer-controlled robot to perform tasks commonly associated
with intelligent beings’ [1]. The principal test of AI is predicated on the Turing Test developed in 1950, by
Alan Turing. This is the test of a machine's ability to exhibit intelligent behavior equivalent to, or
indistinguishable from, that of a human [2]. The current applications of AI range from speech recognition,

DOI: 10.14738/tmlai.83.8219
Publication Date: 06th June, 2020
URL: http://dx.doi.org/10.14738/tmlai.83.8219
 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
finance, avionics, navigation, gaming, robotics and medicine, for example. This paper has been composed
to give an overview of such applications focusing on cryptography, while referencing more technically
oriented papers that are relevant to the material. In particular, the paper focuses on the application of
Evolutionary Computing and Artificial Neural Networks for generating unique and unclonable non-linear
ciphers using real-world noise sources.

1.1 Machine Learning

An important sub-category of AI is Machine Learning (ML). ML is mostly associated with the problem of
pattern recognition. This is where a complex dataset of possibly irregular patterns in a signal or an image,
for example, is required to be categorized into common features and/or segments which can then be
classified in some pre-determined way. These classifications are typically associated with statistical
measures computed from a signal and statistical and/or geometric metrics for an image. If a cluster of
such metrics into a specific numerical range is sufficiently different to be correlated with known features
in the data, then a decision can be made through application of a threshold in order to design a decision-
making criterion. The problem is often how to find an optimum threshold to do this, such that the
accuracy of the decision taken is optimal, the optimum threshold value being subject to a confidence
interval. By making the threshold adapt to the demarcation of certain input metrics in terms of their
known accuracy, quantity and other prior information, the logic of the decision-making process can be
made ‘softer’ in terms of its tolerance to the data. This is the basis for implementing so called ‘Fuzzy Logic
Systems’ which provide the foundations for the development and implementation of Artificial Neural
Networks (ANNs). An ANN typically increases the accuracy of the decisions associated with the
classification of a pattern than can be obtained through conventional data categorization (based on a
logical and/or fuzzy logical quantification). The following section considers the basis for this.

1.2 Artificial Neural Networks

A simple ANN is based on inputting a class of metrics that are taken to be a composite representation of
the data, metrics that are computed by processing the data to form a so called ‘feature vector’. A weight
𝑤𝑤 (with a specific floating-point value) is then applied to determine the significance of each component
of this vector (by multiplying each component of the vector with its weight) and the weighted components
added together to produce a single output value. By changing the values of these weights, the ANN
provides an output in relation to a decision-making process. This is an example of a data transfer process
which transforms multiple inputs into a single output. It can be replicated to produce a number of identical
or different outputs by inputting an identical set of weighted feature vectors or changing the components
of the vector that are input. This generates a ‘single layer’ of outputs that can be taken to be inputs to
some ‘hidden layer’ where the weights are computed and changed again. By replicating this process 𝑛𝑛
times, we can develop a multi-layer network composed of 𝑛𝑛 layers. Figure 1 depicts an example of a
simple, single hidden-layer ANN. In this example, the hidden layer is composed of 4 nodes. The
computations involved in the transfer of information from the hidden layer to the output layer are taken
to include a change in the numerical value to all or some of the weights that are applied. This is an example
of a ‘feed-forward network’ where the ‘information’ flows in one direction only, i.e. forward, from the
input nodes, through the hidden nodes and on to the output nodes. There are no cycles or feed-back loops
in the network. Thus, if the three input nodes given in Figure 1 have numerical values (𝑥𝑥1 , 𝑥𝑥2 , 𝑥𝑥3 ), say,

URL: http://dx.doi.org/10.14738/tmlai.83.8219 22
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

then the inputs to the 4 nodes of the hidden layer are (from top to bottom in Figure 1) given by the
elements of the vector:
(𝑤𝑤11 𝑥𝑥1 + 𝑤𝑤12 𝑥𝑥2 + 𝑤𝑤13 𝑥𝑥3 , 𝑤𝑤21 𝑥𝑥1 + 𝑤𝑤22 𝑥𝑥2 + 𝑤𝑤23 𝑥𝑥3 , 𝑤𝑤31 𝑥𝑥1 + 𝑤𝑤32 𝑥𝑥2 + 𝑤𝑤33 𝑥𝑥3 , 𝑤𝑤41 𝑥𝑥1 + 𝑤𝑤42 𝑥𝑥2 + 𝑤𝑤43 𝑥𝑥3 )

where 𝑤𝑤𝑖𝑖𝑖𝑖 , 𝑖𝑖 = 1,2,3, 4; 𝑗𝑗 = 1,2,3 are the 12 weights required to modify the 3 inputs from the input layer
as they are fed into the 4 nodes of the hidden layer (indicated graphically by the arrows given in Figure 1).
From the description above, and, with reference to Figure 1, it is clear that there are four issues that need
to be considered in the design of any such network of input-outputs: (i) What should the size of the feature
vector be, thereby specifying the number nodes associated with the input layer?; (ii) how many outputs
are required in the output Layer?; (iii) how many hidden layers should be used and how many nodes
should each layer contain?; (iv) how can we automate the process by which the weights are initialized
and/or then adjusted? In regard to point (iv), it is necessary to design an algorithm for automation of the
adjustment of the weights subject to knowledge of the expected output(s) in the output layer. This
requires training data to be provided so that the weights can be updated by comparing the outputs they
provide with ‘target’ data.

Figure 1. Example of a simple neural network consisting of an input layer with three inputs or nodes (i.e. a
three-component feature vector) a single 4-node hidden layer and an output layer consisting of two output
nodes [3].

There are a number of different, but closely related methods (algorithms) to train an ANN by adjusting
the weights until an output to the network for a given input is the same as the training data (within a
specified tolerance). The original and arguably the most common approach, is to apply the back-
propagation algorithm [4] for application to feed forward networks. This is based on the Gradient
Descent Method; an iterative optimization algorithm for finding a local minimum of a differentiable
function. By computing the gradient of a loss function with respect to all the weights (taken to be a
discrete vector) and iterating the process, the weights can be adjusted to generate an output that is close
to the training data. The algorithm has a synergy to the application of a least squares method for solving
a system of linear equations (especially over-determined systems) subject to minimizing an error function
and is an example of dynamic programming. Application of the chain rule is used to compute the gradient
of the loss function with respect to each weight and a threshold function applied which determines
whether or not the value from one node should ‘flow’ further (i.e. continue on to the next node(s) in the
next layer).

Copyright © Society for Science and Education United Kingdom 23

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
There are a wide range of AI software and systems available for the development of AI applications, e.g.
[5] and [6], respectively. MathWorks MATLAB, for example, provides specialist Toolboxes including
capabilities that integrate AI into the complete workflow for developing fully engineered systems [7] and
machine learning with supervised and unsupervised learning [8]. The development of AI systems using
Python programming is becoming increasingly common [9]. However, whatever system and/or
programming language or toolbox is used, the design of a network is fundamental and must be tailored
to optimize the application of AI using an ANN for a specific solution. Of specific importance, is both the
quality and quantity of the training data. This determines the accuracy of the weights that, along with the
design of the network, can be thought of as the ‘keys’ to the future operation of the network. Thus, there
is an obvious application in cryptography where the weights are analogous to the key(s) and the network
design, to the encryption algorithm for generating an output cipher that can then be used to encrypt
plaintext data.

1.3 Artificial Neural Networks, Data Processing and Deep Learning

As discussed in Section 1.2, it is typical to first of all process the data to generate a feature vector
containing metrics that are taken to be a good representation of the essential characteristics of the data,
a digital signal, for example. In this way, the number of nodes in the input layer become relatively few
compared to the original data, i.e. the length of the digital signal. This is important in regard to utilizing
the inevitable limited computational power available to ‘drive’ an ANN in order to produce an efficient
decision-making process (e.g. the computational time required). However, in some cases, it is very difficult
to define, in a fully quantitative sense, the elements of the feature vector which are good (unique and
unambiguous) characteristics of the data. This problem is often overcome by investigating new properties
of the data based on novel analysis methods. For example, texture in an image can be quantified using
the principles of fractal geometry and computing metrics such as the Fractal Dimension and multi-fractal
parameters. This allows more impressionable features in an image (or the image as a whole), for example,
described by the rather elusive term ‘texture’, to be quantified through Fractal Geometry in Digital
Imaging [10] and [11].
The determination of what feature vector should be used, and the elements it contains, determines the
size of the feature vector which should ideally be much smaller than the original data from which it has
been constructed using a number of different signal or image processing algorithms to output specific
metrics. There is a optimality issue that needs to be considered, which is the processing time required to
compute the feature vector relative to the time required to train the ANN. However, in recent years, the
computational power available through hardware improvements has increased to such a degree, that
instead of processing signals to yield a composite feature vector, this process is ignored and the raw data
(i.e. of the original digital signal) is used directly to ‘feed’ the input layer of an ANN. Coupled with many
layers consisting of a large number of nodes, this approach significantly increases the complexity of the
ANN but reduces that need to decide what signal processing algorithms are required to generate a smaller
feature vector a priori.
This is the principle associated with a more recent approach to AI known as Deep Learning in which prior
signal processing methods are effectively abandoned and high data throughput ANNs designed to process
the original data alone. In this context, Figure 2 illustrates the difference between a conventional single
hidden layer or ‘shallow’ ANN and a deep network consisting of three hidden layers.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 24
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

The principal issue that is driving the deep learning approach is the increase in computing power and
efficiency. For example, Google’s Inception-V3 is based on using a network consisting of 49 layers with
more than 20 million interconnections [13]. But this is of little value unless there is a sufficient amount of
training data to compute the millions of weights now required. However, with the increase in
computational power coupled with the exponential growth in the Internet of Things (IoT), it is becoming
much easier to obtain online access to the training data required. In this sense, deep ANN’s are possible
because of the growth in the Big Data Society that is now prevalent as well as the computational
performance available. The growth in computing power, which is allowing a deep learning paradigm to
evolve, is primarily industry driven by companies such as Facebook, Google, and, by national security
services where issues such as face recognition and track and trace, for example, are primary concerns.
This is possible because super computers from some ten years ago, consisting of thousands of processor
units and accommodated in buildings with a large floor space can now be condensed into single Graphical
Processing Units (GPUs) required for high resolution gaming, for example.

Figure 2. Examples of a single layer ‘shallow’ neural network (left) that has a 5 -element feature vector with
a single layer and a deep neural network consisting of 4 hidden layers, both networks consisting of a 4-node
output layer [12].

Irrespective of the computational power available, ANNs require real world data to operate in real world
environments. Although this necessity is catered for by the development of access to big data, in some
cases, certain data may not be available or is missing from existing data fields. In this case, ANN’s can also
be of value by training them to create the missing data using other related (real world) data that is
available. Thus, not only is AI benefiting from the big data society but can be active in the development,
growth and diversity of the big data society (by creating even more data).
Coupled with the training data, the design of an ANN can vary considerably (i.e. the number of layers, the
nodes per layer and homogeneity or otherwise of the connections provided etc.) as can the details of the
algorithms designed to compute the weights. This includes diffusing the values of the weights over a layer
and/or a partition through the layers to a specific depth by applying various spatially invariant filters using
the (discrete) convolution operation. These Convolutional Neural Networks (CNNs) were first introduced
in the early 1980’s. They are a class of deep learning ANNs developed for and applied mainly in the fields
of digital image processing and computer vision [14]. In this context, they are replacing many digital image
processing algorithms originally developed for pattern recognition using (among other techniques)
convolutional filters to segment images into regions of similarity and/or discontinuity (edge detection, for
example) [15]. This approach has and continues to be used to develop a set of deterministic, geometric

Copyright © Society for Science and Education United Kingdom 25

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
and textural metrics that are traditionally used to design a fuzzy logic decision engine or to train ‘shallow
learning’ ANNs with relatively few inputs and hidden layers.
The application of deep learning solutions to solve complex pattern recognition problems is key to the
future development of many technologies and programming environment, functions and systems, all of
which are in the process of rapid development. For example, MATLAB have recently developed a new
toolbox for solving problems using deep learning [16]. The application of deep learning to cryptography
lies beyond the scope of this paper. However, some of the approaches considered in this paper naturally
lend themselves to improvements in performance using a deep learning approach. This is especially so in
regard to replacing conventional cryptanalysis methods which, in regard to some of the developments
presented in this paper, are becoming increasingly redundant [17]. The following section introduces the
role ANNs in Cryptography.

1.4 Artificial Neural Networks and Cryptography

One of the principal features of AI is its ability to recognize patterns within complex data. Cryptography
relies on maximizing the diffusion and confusion associated with the conversion of plaintext into
ciphertext. Ideally, the encrypted data – the ciphertext – should be entirely devoid of any type of pattern.
It is required that the ciphertext is a complex randomized representation of a plaintext usually associated
with the application of a one-way function which has no inverse solution. This provides the potential for
the application of AI in regard to the generation of ciphers, their classification (i.e. their cryptographic
strength) and the analysis or cryptanalysis of encrypted data.
Since cryptography relies on the conversion of plaintext into a ciphertext based on the generation of a
random number field (the cipher), suppose we input a few initial random numbers to seed a network and
then train it using a real source of noise, allowing the weights to be adjusted so that the network outputs
an reasonably accurate simulation of the target random data field. In this case, providing the design of
the network is known to two communicating entities, Alice and Bob, the initial random numbers and the
weights represent the key(s) necessary to reproduce the random number field that can be used to encrypt
and then decrypt the data (subject to both the initial random numbers and weights being known to Alice
and Bob through application of a key exchange protocol).
The network design (including the computed weights) is equivalent to the encryption/decryption
algorithm. In this way, a low number of random elements in the input feature vector can generate a
random number field that constitutes the cipher. Such an approach is discussed later on in the paper and
compared to another solution to the problem of developing an algorithm for generating ciphers. This is
based on the application of a machine learning technique called Evolutionary Computing which can be
used to provide a nonlinear equation that, upon iteration, can produce a random number field, the initial
condition of the iterator being the (conventional) key.
Just as AI can be used to generate a cipher, the ability for AI to recognize complex patterns lends itself to
undertaking the task of recognizing patterns in a ciphertext and thereby looking for signatures that
determine a proximity or ‘point of attack’ where there are weaknesses in the randomized character of the
encrypted data. Given that a strong encryption engine should not generate such weaknesses, AI provides
a complementary approach to testing the strength of an encryption method, including those methods
that are based on the use of AI for generating the encrypted data. Thus, AI has roles to play in association

URL: http://dx.doi.org/10.14738/tmlai.83.8219 26
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

with both data encryption and cryptanalysis. In the latter case, this can be extended to problems
associated with information hiding when it is required to detect the signature of a plaintext hidden in
another plaintext (Steganalysis) and the signature of a ciphertext hidden in a plaintext
(Steganocryptanalysis). In the following section, we briefly review the role of ANNs in Cyber Security.

1.5 Artificial Neural Networks and Cyber Security

Cyber security involves a range of techniques designed to protect the creation, processing, storage and
transmission of data over an open network, such as the Internet. A specific and well-known example of
this is the infection of files by a virus designed to covertly infiltrate a single computer or a network of
computers. Computer viruses have become complex and operate in a stealth mode to avoid detection.
New viruses are created each and every day. However, most of these supposedly ‘new’ viruses are not
totally new but are often variations (some of them relative minor variations) on the theme of their
predecessors. Many of these viruses just change their form and signatures to avoid detection but their
operation and the way they infect files and systems is still the same. This provides the potential for training
ANNs on a range of known viruses under the assumption that their digital signatures can be recognized
by inspecting future data streams and new files, e.g. [18] and [19].

1.6 About this Paper

Having provided an overview of AI and its applications to Cryptography and Cyber Security in this Section,
the paper now provides some specific studies based on previous research already conducted by the
authors’ coupled with some of their current research themes. Studies are presented which include: (i) the
application of Evolutionary Computing to generate unique and unclonable ciphers; (ii) a machine learning
strategy for detecting the randomness (or otherwise) of binary streams designed to investigate the
potential to attack a ciphertext stream. In the former case, an overview of the authors’ current research
is given in regard to authenticating documents using a Smartphone as discussed in Section 4.

2 Cryptography using Evolutionary Computing

In Patrick Mahon’s secret history of ‘Hut 8’ – the Naval Section at Bletchley Park (Station X) from 1941-
1945 - it is stated that [4]: ‘The continuity of breaking Enigma ciphers was undoubtedly an essential factor
in our success and it does appear to be true to say that, if a key has been broken regularly for a long time
in the past, it is likely to continue to be broken in the future, provided that no major change in the method
of encryption takes place’. This statement is in regard to the famous Enigma encryption machine used by
German armed forces from the mid-1930s until 1945, the underlined component of this statement
relating to the design of the enigma machine and not its key settings.
The design of the enigma machine is an example of what today we would call an encryption algorithm
whose output is ultra-sensitive to the input key. Today, such encryption algorithms often manifest
themselves in terms of just a few lines of source code. They are typically and traditionally based on the
application of modular arithmetic coupled with certain mathematical ‘tricks’ associated with the
properties of prime numbers, for example. This includes the famous Rivest, Shamir, Adleman (RSA)
algorithm which is arguably the foundation stone for the development of public-private key cryptography
whose security relies on the difficulty of factoring large semi-prime numbers into two component prime
numbers. Such a factorization has been computationally impossible with conventional digital computing
(even with super computers) using procedural algorithms such as Shor’s algorithm. However, with the

Copyright © Society for Science and Education United Kingdom 27

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
development of quantum computers which allow this algorithm to be implemented efficiently, prime
number-based encryption is set to become increasingly obsolete.
Irrespective of the encryption algorithm available (in the past, present or future), there is a conventional
rule called the Kerckhoff-Shannon principle which states that a crypto-system (including the algorithm
with pre- and post-data processing) should be secure even if everything about the system, except the
key(s) is public knowledge. This is stated more succinctly by Claude Shannon as ‘the enemy knows the
system’. Certainly, it is important that an encryption algorithm can be scrutinized by others in order to
confirm (publicly or otherwise) its cryptographic strength. But this is correlated to a certain extent with
the relative difficulty of producing new algorithms by hand. AI has overcome such difficulties, and, it is
now possible, as is discussed in this section, to produce different encryption algorithms as easily (in
principle) as it is to produce different keys. In this context, the encryption algorithm becomes the key, but
like a conventional key, still needs to be exchanged between sender and receiver. This allows us to break
with the conventional Kerckhoff-Shannon principle. The approach is to apply real world data, in this case,
real world noise, which is used to either train an ANN to simulate the noise as briefly discussed in Section
1.4, or, to output a nonlinear function that approximates the noise and can then be cast as an iterator
where the initial condition is the (conventional) key. In regard to the latter case, we present an approach
based on the application of a machine learning method called Evolutionary Computing.

2.1 Background to the Case

Consider an encryption method that is based on the Gilbert Vernam Cipher [21] which is in fact, a perfectly
secure method of encrypting data for one-to-one communications provided a secure method of key
exchange is available. The Vernam cipher is a substitution cipher based on generating an array of random
numbers to form a vector 𝒙𝒙 = (𝑥𝑥1 , 𝑥𝑥2 , … , 𝑥𝑥𝑁𝑁 ) – the ‘cipher’. This vector represents a digital signal that is
taken to be a stochastic field - a purely noise driven signal. The problem is how to generate an algorithm
that can be executed on a digital computer to output such a vector that is suitable for encryption using
the Vernam cipher or otherwise.
Suppose that some plaintext is written in terms of a set of numbers (using the ASCII, for example), thereby
constructing a plaintext vector 𝒑𝒑. The plaintext is typically taken to be the text associated with a natural
language, the plaintext vector consisting of decimal integer numbers conforming to the ASCII for the
natural language, but, in principle, any code can be used. However, in a more general context, the
plaintext vector could consist of elements representing any signal or image, for example. In the latter case,
the elements would typically be decimal integers in the range 0-255 for an 8-bit grey level image, but in
the former case, the elements may be floating point numbers as can the elements of the cipher. Either
way, for a substitution cipher, we can generate the ciphertext by simply adding the two vectors together
to generate the ciphertext 𝒄𝒄 = 𝒙𝒙 + 𝒑𝒑. In the case when the plaintext is natural language based text data,
coded using 7-bit ASCII and composed of 𝑁𝑁 elements (when the cipher is also composed of 𝑁𝑁 integer
values), we can construct the ciphertext as given in Equation (1) below (where mod denotes the modulo
operation):
𝑐𝑐𝑘𝑘 = (𝑥𝑥𝑘𝑘 + 𝑝𝑝𝑘𝑘 ) mod (127), 𝑘𝑘 = 0,1, 2, … , 𝑁𝑁 − 1 (1)
The plaintext is then recovered from the decryption equation 𝑝𝑝𝑘𝑘 = (𝑐𝑐𝑘𝑘 − 𝑥𝑥𝑘𝑘 ) mod (127) , which
requires 𝒙𝒙 to be known of course. However, another way of implementing the method of encryption is to

URL: http://dx.doi.org/10.14738/tmlai.83.8219 28
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

write the integer plaintext and cipher vectors as binary strings (using ASCII, for example). In binary space,
the ciphertext is then given by 𝒄𝒄 = 𝒙𝒙 ⊕ 𝒑𝒑 where ⊕ denotes the binary exclusive OR (XOR) operator, the
decrypt being given by 𝒑𝒑 = 𝒙𝒙 ⊕ 𝒄𝒄. In this case, the vector notation used to denote the binary space data
𝒄𝒄, 𝒙𝒙 and 𝒑𝒑 denotes binary strings that are of a finite length L>N where typically, L>>N for a standard
plaintext massage.
Whatever the method of encryption that is implemented, a principal issue is how to design an algorithm
or a class of algorithms that output a cipher with properties that are consistent with strong encryption.
These properties include ensuring that 𝒙𝒙 is statistically unbiased so that the histogram of the cipher is
uniformly distributed and equally so, has a power spectral density function that is uniform. Most cipher
generating algorithms are based on iterations in which the initial value is the key. They produce pseudo
random number streams for which certain critical conditions are required to be met. These conditions
include: (i) ensuring that the algorithm generates random numbers that are equally and uniformly
distributed irrespective of the key that is used; (ii) given that the cycle length of any finite state
computations is itself finite, is the characteristic cycle length of the iteration longer than the length of the
plaintexts that are to be used, thereby avoiding patterns in the random number stream that are correlated
cyclically.
There are numerous cipher-generating algorithms based on the design of Pseudo Random Number
Generators (PRNGs) that have been developed. They tend to fall into three classes which generate: (i)
decimal integer random number streams; (ii) floating point random number streams; (iii) random binary
streams. The traditional focus has been on the computation of integer streams because of the
computational efficiency associated with integer arithmetic. This has typically involved the coupling of
modular arithmetic with prime numbers which is why prime number-based cryptography has evolved in
the way that it has. By way of an example, the Blum Blum Shub (BBS) cipher [22] is a PRNG first proposed
in 1986 that is based on the iteration (for 𝑘𝑘 = 0, 1, 2, . . . , 𝑁𝑁 − 1)
𝑥𝑥𝑘𝑘+1 = 𝑥𝑥𝑘𝑘2 mod (𝑀𝑀), 𝑀𝑀 = 𝑝𝑝𝑝𝑝 (2)
where p and q are prime numbers and 𝑥𝑥0 is the key (the initial condition or ‘seed’) which is a co-prime
to M meaning that p and q are not factors of 𝑥𝑥0 and not 1 or 0. This algorithm operates on, and outputs a
string of pseudo random values that are decimal integers.
The design of cipher generating algorithms that output decimal integers evolved in parallel with the
limited computing power available primarily, the limited processing time associated with floating point
array processing. With the more recent development of fast floating-point processors and co-processors
and specialist real-time digital signal processors, floating-point cipher generation has been able to exploit
the study of chaos to produce a new class of chaotic iterators that are based on non-linear iterations and
require high precision floating-point arithmetic to be performed.
An example of such an iterator is the Vurhulst cipher given by
𝑥𝑥𝑘𝑘+1 = 4𝑟𝑟𝑥𝑥𝑘𝑘 (1 − 𝑥𝑥𝑘𝑘 ), 𝑟𝑟 ∈ (0,1) (3)
which has a certain synergy with the BBS PRNG given that both are quadratic iterators. In this case, the
initial condition 𝑥𝑥0 is a floating-point number between 0 and 1. However, this iteration only provides full
chaos when r=1 which is prohibitive given that 𝑥𝑥𝑘𝑘 converges and then bifurcates multiple times as r

Copyright © Society for Science and Education United Kingdom 29

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
approaches 1. For this reason, a modification to Equation (3) can be introduced by considering the
iteration [23]
1 𝑟𝑟
𝑥𝑥𝑘𝑘+1 = 4(1 + 𝑟𝑟) �1 + � 𝑥𝑥𝑘𝑘 (1 − 𝑥𝑥𝑘𝑘 )𝑟𝑟 𝑟𝑟 ∈ (0,1)
𝑟𝑟
A range of values of 𝑟𝑟 can then be used in addition to an initial value 𝑥𝑥0 ∈ (0,1) which extends the
iterator to a two-parameter algorithm.
In general, we can consider a generic iterative cipher to be of the form
𝑥𝑥𝑘𝑘+1 = 𝑓𝑓(𝑥𝑥𝑘𝑘 ), 𝑥𝑥0 ∈ (0,1) (4)
where 𝑓𝑓 is some nonlinear function which may include a parameter or set of parameters whose numerical
values (or range of values) need to be establish a priori to provide a chaotic cipher 𝒙𝒙. In principle, one
could attempt to generate a data base of many such non-linear functions. These functions might be
modifications of known chaotic maps developed to investigate specific physical (feedback) models of
chaos or simply invented. There are, however, a number of issues that need to be appreciated in order to
implement such an approach. These issues are considered below.
2.1.1 One-way functions
In the application of Equation (4) to cryptography, it is assumed by default that 𝑓𝑓 is a one-way function
which means that 𝑥𝑥0 (the key) cannot be derived from a knowledge 𝑥𝑥𝑘𝑘 . However, for certain functions
and under certain conditions, this is not the case. For example, with 𝑟𝑟 = 1, Equation (3) has the analytical
solution [24]

𝑥𝑥𝑘𝑘 = sin2 (2𝑘𝑘 sin−1 �𝑥𝑥0)

Thus, for 𝑘𝑘 = 1, it is possible to invert this equation to obtain the key 𝑥𝑥0 meaning that the key can be
obtained from knowledge of the first iteration (or higher order iterations) which is clearly not acceptable
in the design of a cipher generating algorithm. This result has a synergy with Equation (2) in so much as
this equation can be written in the form
𝑘𝑘 mod 𝐶𝐶(𝑀𝑀)
𝑥𝑥𝑘𝑘 = �𝑥𝑥02 � mod 𝑀𝑀

where 𝐶𝐶 is the Carmichael function [25]. Thus, in order to use an iteration of the type given by Equation
(4), the nonlinear function must be proved to be a one-way function and not to have an equivalent
analytical solution that is invertible.
2.1.2 Equal Probability of States
What-ever nonlinear function is applied, modified or invented, the distribution of 𝒙𝒙 is rarely uniform and
the vector must be post-processed to generate such a distribution. This can be done by windowing all
values of the vector that conform to a uniform distribution. However, this wastes data and processing
time required to generate it, given that many computed floating-point values may have to be discarded.
This is an important issue because in assuming the existence of one-way functions, there can exist
probability distributions, which are not uniform and are not even statistically close to a uniform
distribution, but are nevertheless, computationally indistinguishable from real-world chaos [26]. Hence,
checking for equal probability of the states is fundamental. One way of enforcing this requirement is to

URL: http://dx.doi.org/10.14738/tmlai.83.8219 30
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

implement a partitioning strategy to output a random binary string as illustrated in Figure 3, which shows
the histogram for Equation (3) and is uniformly distributed for 𝑥𝑥𝑘𝑘 ∈ [0.3,0.7] (approximately). A binary
string can therefore be created by outputting a ‘0’ when 𝑥𝑥𝑘𝑘 ∈ [0.3,0.5] and a ‘1’ when 𝑥𝑥𝑘𝑘 ∈ (0.5,0.7] or visa-
versa. In this way, a so-called maximum entropy bit stream cipher is obtained in which the encryption
process is undertaken using a standard XOR operation.

Figure 3. Histogram of the output for a Vurhulst process given by Equation (3) and the partitioning strategy
used to generate a maximum entropy binary cipher [24].

Further examples of such ciphers include the functions or maps given in Table 1 which specify the form of
the function given in Equation (4). These are examples of stream ciphers obtained by inputting a key
𝑥𝑥0 ∈ (0,1), optimizing their output (through adjusting the value of 𝑟𝑟) to produce a chaotic number-stream
and post processing this output to generate a stream that is uniformly distributed [27].
Table 1. Examples of chaotic maps for the generation of stream ciphers.

Name of Map Function

Generalized Tent Map 𝑟𝑟(1 − |2𝑥𝑥 − 1|𝛼𝛼 ), 𝛼𝛼 > 0
Quadratic Feedback Map 𝑟𝑟𝑟𝑟[1 − 𝑥𝑥(1+𝑥𝑥 2)]
Generalized Sine Map |sin(π𝑟𝑟𝑥𝑥 𝛼𝛼 )|, 𝛼𝛼 > 0
Tangent Feedback Map 𝑟𝑟𝑟𝑟[1 −tan(1/2𝑥𝑥)]
Logarithmic Feedback Map 𝑟𝑟𝑥𝑥[1 − 𝑥𝑥log(1+𝑥𝑥)]

Not all invented or otherwise maps are suitable from a probabilistic point view even though they may
exhibit chaos. For example, the map 𝑓𝑓(𝑥𝑥) = 𝑟𝑟|1 − tan [sin(𝑥𝑥)]|, 𝑟𝑟 = 1.5 has a highly non-uniform
distribution over all probability states and can therefore not be conditioned to produce a maximum
entropy cipher using the partitioning strategy illustrated in Figure 3, for example.

2.2 The Lyapunov Exponent

In addition to ensuring equal probability states, it is also necessary to make sure that the nonlinear
function is characterized by an iteration in which the Lyapunov exponent is positive, thereby guaranteeing
that the iteration is not convergent but chaotic [24]. For a long sequence of numbers 𝑥𝑥1 , 𝑥𝑥2 , … , 𝑥𝑥𝐾𝐾 , a useful
measure of the Lyapunov exponent is given by
1
λ(𝑥𝑥0 ) = ∑𝐾𝐾
𝑘𝑘=1 log| 𝑓𝑓′ (𝑥𝑥𝑘𝑘 ) | (5)
𝐾𝐾

We require that λ > 1 for all 𝑥𝑥0 ∈ (0,1). A high, but strictly positive value of the Lyapunov exponent is
preferable because the iteration function it is taken to characterize will then generate chaotic trajectories
within a few iterations. Thus, ideally what we require is a nonlinear function for which λ ≫ 1 for all

Copyright © Society for Science and Education United Kingdom 31

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
𝑥𝑥0 ∈ (0,1). In this context, the nonlinear function must also generate an iteration with a long cycle length.
In practice, the actual values of the Lyapunov exponent and the cycle length are relative to some known
iteration that has been categorized as being cryptographically strong a priori with regard of the measures
discussed. These issues are difficult to determine analytically and require numerical tests to be
undertaken which includes measuring the processing time which cannot be excessive.
2.2.1 Structural Stability
Numerical tests do not guarantee the diffusive properties of a cipher, namely, that the PRNG is structurally
stable. Ideally, we require an iteration that has (almost) the same cycle length and Lyapunov exponent for
all initial conditions. Most of the known pseudo-chaotic systems do not possess this property and there is
no rigorous analytical method, as yet known, for assessing this property. This is an important problem
because without solving it, it is not possible to guarantee that a crypto system based on a deterministic
chaotic algorithm or set of algorithms will always produce uncorrelated strings for any and all keys,
irrespective of their floating-point accuracy.
2.2.2 Computational Unpredictability
Another issue is the determination of unpredictability for chaotic systems, i.e. what properties of a chaotic
system grantee its computational unpredictability. This issue is coupled with the problem of algorithmic
complexity which cannot be commuted; there is no universal solution for simplifying programs and for
proving that the length is minimal. We cannot apply this definition directly to compare the complexity of
cryptographic sequences or algorithms. Nevertheless, the theoretical applications are very important. In
particular, the Kolmogorov complexity provides a unified approach to the problem of data compressibility
[28]. Subject to these important and, as yet, unresolved issues, the applications of ‘digital chaos’ can yield
commercially realizable products but not necessarily products that are scalable in terms of commercial
capacity.
Given the issues discussed above in regard to inventing a suitable nonlinear function, it would be
beneficial if an approach could be developed where such a function could be evolved from real-world
stochastic data, on the understanding that the function would provide only an approximation to the data.
To accomplish this, we resort to the use of Evolutionary Computing which is a form of machine learning.

2.3 An Introduction to Evolutionary Computing

In Computer Science, evolutionary computation is known as a family of algorithms for global optimization
that are inspired by biological evolution [29]. In practical terms, they are a family of population-based trial
and error problem solvers, with a metaheuristic or stochastic optimization character [30], i.e. an initial set
of candidate solutions is generated and iteratively updated. Each new generation is produced by
stochastically removing less desired solutions, and introducing small random changes. We can say
therefore, that evolutionary computational techniques have the capability to produce highly optimized
solutions across a wide range of problems. The principles of evolutionary strategies where first introduced
in the 1960’s [31]. Evolutionary programming was later introduced in the 1990’s [32] and while these
areas developed separately, by the late 1990’s, they were unified as different representatives or ‘dialects’
of one technology, namely, Evolutionary Computing (EC) [33]. Around the same time, a fourth stream,
following the same general concepts had also emerged known as ‘genetic programming’ [34]. Since then,
nature-inspired algorithms have become an increasingly significant part of evolutionary computation.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 32
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

Many aspects of EC are stochastic, but the starting point of candidate solutions can be either deterministic
or stochastic. In EC, genetic algorithms deliver methods to model biological systems that are linked to the
theory of dynamical systems, since they are used to predict the future states of the system.
Within the field of EC itself, a software system called ‘Eureqa’ was one of the first such systems to be
developed by the Cornell Creative Machine Laboratory (Cornell University, USA) and then commercialized
by Nutonian Inc. [35]. The underlying principle is to use genetic programming to generate dynamic
equations, each of which provides an increasingly better ‘fitness function’ to model a given dataset. If this
dataset is complex, such as a noise field, the system iteratively generates a sequence of non-linear
functions to approximate the input signals [36]. Thus, Eureqa is a modeling engine predicated on machine
learning, using evolutionary searches to determine an equation that best represents a given data set.
Eureqa claims to ‘automate the heavy lifting inherent in analytics and data science’ [37]. The system
automatically discovers analytical models requiring almost no human intervention and randomly creates
equations via sequences of mathematical building blocks based on a combination of common functions.
It is therefore suitable for generating non-linear functions by seeding the system with data from real world
noise sources. The functions can then be iterated as required, to produce a key seeded pseudo random
number sequence.
Evolutionary Computing is associated with the field of Computational Intelligence, and like AI, involves
the process of continuous optimization. AI aims, through iterative processes, to compute a set of optimal
weights that determine the flow of information (the amplitude of a signal at a given node) through a
network that simulates a simple output subject to a complex input. In this sense, an ANN simulates a high
entropy input with the aim of transforming the result into a low entropy output. However, this process
can be reversed to generate a high entropy output from a low entropy input. In this sense, ANNs can be
used to generate ciphers by simulating natural noise once it has been trained to do so.

2.4 Stream Cipher Generation using Evolutionary Computing

The principal purpose of using EC and ANNs to produce stream ciphers is that the example maps given in
Table 1 must derived, modified or just invented subject to an application of the conditions required for
them to be compatible with application in cryptography. EC has the potential to do this automatically
using real world noise to initiate the evolutionary process. While an ANN approach to generating ciphers
is of value in special cases, it does not provide the same flexibility in terms of designing PRNGs using
iterated (nonlinear) functions. To do this, an evolutionary algorithms approach is required in which a
population-based, stochastic search engine is required that mimics natural selection. Due to their ability
to find excellent solutions for difficult and dynamic problems within acceptable computational time,
evolutionary algorithms have therefore attracted interest from many areas of science and engineering.
The application of evolutionary computed algorithms to cryptology was first consider in 2013 [38] using
Eureqa [37]. This system is used to iteratively develop a nonlinear function to describe complex input
signals usually associated with experimental data of a chaotic system. If genuine random (delta
uncorrelated) noise is input into the system, then from a theoretical point of view, no nonlinear function
will be found on an evolutionary basis that models this noise perfectly. Thus, inputting natural noise is a
way of ‘cheating’ the system to ‘force’ it to provide an approximation to the noise that may be suitable
(on an iterative basis) as a PRNG. This suitability is based on a set of tests including an evaluation of the
Lyapunov Exponent, checking that the cipher stream is uniformly distributed, as is the Power Spectral

Copyright © Society for Science and Education United Kingdom 33

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
Density Function (PSDF) and so on, as outlined in Diagram 1 (where the ticks indicate that a given test has
been passed). By comparison, Diagram 2 shows an equivalent schematic for using an ANN to produce a
stream cipher. Comparing the two approaches, it is clear that the weights and the ANN design are
equivalent to an evolved function using EC.
In either case, and, in addition to any function being tested against the checks given in Diagrams 1 and 2,
the National Institute of Standards (NIST), Computer Security Resource Center, provides a Cryptographic
Algorithms Validation Program (CAVP) [39] which includes testing new PRNGs [40]. These NIST tests are
an international standard for validating the cryptographic strength of a cipher and are mandatory in the
use of a cryptographic algorithm.

Diagram 1. Schematic of the processes for evolving a stream cipher using EC.

Diagram 2. Schematic of the processes for generating a stream cipher using an ANN.

2.4.1 Real World Noise Sources

There are a number of real-world noise sources that can be used to input into the systems whose
schematics are given in Diagrams 1 and 2. For example, Random.org is a free internet resource that
provides true random number streams [41]. In this case, the random data is derived from atmospheric
noise generated by radio emissions due to lightening; there are approximately 100 lightning strikes to
earth per second. Another example is the quantum mechanical noise generated using a reverse biased

URL: http://dx.doi.org/10.14738/tmlai.83.8219 34
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

semiconductor junction. This can be provided in the form of an external USB interface manufactured and
supplied by Araneus Information Systems in Finland, for example. Their Alea II is a compact true random
number generator, also known as a hardware random number generator, non-deterministic random bit
generator, or entropy source [42].
2.4.2 Evolved Cipher Generation Examples
Given that this paper is a publication composed for the Transactions of a Society for Science and
Education, the authors have chosen to present examples based on previous and current teaching activities
associated with the Information and Communications Security course (COS 792) given as part of the
Honors program in Computer Science in the Department of Computer Science at the University of Western
Cape (UWC), South Africa. One of the course work assignments is as follows: Read the paper
‘Cryptography using Evolutionary Computing’ by Blackledge et al. 2013 [38]. Using the information and
online references discussed in this paper, develop another example of the cipher given in Section V of this
paper using the same approach and online facilities.
Figure 4 shows a screen shot of Eureqa which generated the map given by the following non-linear
function:
𝑓𝑓(𝑥𝑥) = 0.7529 − 0.4697 sin(0.7529 + sin(sin(sin(−4.334 × 105 cos(𝑥𝑥)))))
− 0.569𝑥𝑥 cos(𝑥𝑥) acos(𝑥𝑥) sin (4.059𝑥𝑥 2 sin (sin (−4.331 × 105 cos(𝑥𝑥)) − 2.212 × 105 cos (𝑥𝑥)))

Figure 4. Screen shot of Eureqa evolving a nonlinear equation to approximate an input vector consisting of
random data obtained from Random.org.

obtained by one of the students of the 2018 class [43]. The figure shows the outputs of the system which
includes the list of equations obtained post 100 iterations, the solution chosen, its goodness of fit, the
correlation coefficient, the maximum error and the mean squared error. The plots given in Figure 4 are
of the normalized solution fit to an input vector consisting of 138 random data elements (top-right) and
the solutions accuracy (Error) plotted against the complexity (lower-right) as the iterations evolve
different equations (of greater complexity but with a lower error). The selected equation had a 50%
Stability and 76.2% Maturity and used 95% of the training data obtained from Random.org [41].
By repeating the process and using Eureqa to generate functions that approximate different random
vectors (obtain using different noise sources or different noise fields from the same source), it is possible
to generate an unlimited number of key dependent stream ciphers by iterating the function obtained

Copyright © Society for Science and Education United Kingdom 35

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
according to Equation (4). Table 2 shows some examples of maps that have been randomly selected from
the assignment submissions produced by students of the Information and Communications Security
course (COS 792) at UWC from 2016-2019. Each map is named after the student that evolved the function
using Eureqa. These are examples of many hundreds of such functions that have been evolved by students
of the course since 2016.
Table 2. Examples of maps evolved using Eureqa by a selection of Honors students at UWC.

Name of Student, Date Function

D J van Roodt, 01-10- acosh(𝑥𝑥 + 𝑥𝑥 3 asinh(𝑥𝑥 − 1)) acosh �𝑥𝑥 − 1
2016
+ 𝑥𝑥 3 �6 + atan�𝑥𝑥asinh(𝑥𝑥 3 )��� 𝑥𝑥 3

G Chibba, 28-11-2017 0.5076 + 0.1094 sin(1.574𝑥𝑥) + 0.1568sin (6.107 +

𝑥𝑥log�0.4619 sin�−0.04284𝑥𝑥sin(−0.03866𝑥𝑥)��)

J Christians, 29-11-2018 44.42 + 21.91sin(3.362 + cos(47.26cos(41.49𝑥𝑥)) − 32.66𝑥𝑥 2 +

20.95cos (4.23 + 53.46 cos(sin(3.362 + cos(47.26 cos(41.49𝑥𝑥)) −
32.66𝑥𝑥 2 ))2 sin(3.362 + cos(47.26 cos(41.49𝑥𝑥)) − 32.66𝑥𝑥 2 )) +
cos (41.49𝑥𝑥)
R Lehata, 30-11-2019 cos(1.691 + 2.216𝑥𝑥) − 0.9704𝑥𝑥cos(1.807 + 2.03𝑥𝑥) −
0.2623cos (1.314 𝑥𝑥sin(𝑥𝑥) cos (cos (1.807 + 2.03𝑥𝑥) −
0.9884 𝑥𝑥cos (1.807 + 2.03𝑥𝑥)) − 5.863 × 105 𝑥𝑥)

2.5 Discussion
While the output maps evolved by Eureqa should be routinely tested against the NIST CAVP (since the
maps are essentially an approximation of real-world noise), it can be expected that they will conform to
the basic tests for randomness. Thus, the distribution will, by-default, be uniformly distributed and there
is no need to apply a partitioning strategy as illustrated in Figure 3, for example, to generate a maximum
entropy cipher. In principle, any evolutionary computed cipher can be iterated according to Equation (4),
so that it can be used a number of times for a different key 𝑥𝑥0 ∈ (0,1).
Another approach is to use an EC ciphers once and once only, in which case, iteration is not required. This
approach over comes the issue of ensuring that the cipher is structurally stable so that the iterator has
(almost) the same cycle and Lyapunov exponents for all initial condition (given that most of the known
pseudo-chaotic systems do not possess this property). However, this assumes that a large data-base of
such ciphers has been created a priori and/or can be created in real time as and when required. But since
each cipher requires a real-world noise sequence to generate a map, it is arguable that instead of using
EC to evolve a map, the original data might be used for encryption instead, which then obviates the need
for EC in the first place. On the other hand, since either the map or the data from which it has been
evolved must be exchanged between Alice and Bob, the exchange of a map will be preferable given that
maps such as those given in Table 2 require less bits to transmit over an unsecured network using a known
key exchange algorithm. In either case, a key exchange algorithm is necessary to exchange the map (once)

URL: http://dx.doi.org/10.14738/tmlai.83.8219 36
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

and then the keys used to generate different initial conditions for each plaintext, or the map alone. In the
latter case, the map becomes equivalent to the key for a known algorithm.
In addition to providing a cipher, the method can also be used to generate encryption keys. This has been
considered in [44] which uses EC to generate keys for encrypting cloud-bound data with Advanced
Encryption Standard (AES-128, 192 and 256), the Data Encryption Standard (DES and 3DES) and a novel
cryptosystem called ‘Cryptor’. One of the problems faced in cryptography is key management. The key
management in this case is addressed using an ANN to learn patterns of the encryption key. Once learnt,
the key is then discarded to thwart an attack on the key. The key is then reconstructed by the ANN for the
purpose of decryption.
2.5.1 Issues on Practical Implementation
An important issue is that EC generated maps can take a significant amount of CPU time to evolve on a
standard personal computer, for example, and once generated, may require a large number of high
precision floating point operations to compute. In general, the longer EC is left to evolve a function, the
more complex the output map becomes. However, this issue is essentially a hardware development
concern, and, like the rise of deep learning using complex and computationally intensive ANNs (as
discussed in Section 1.3), provides a way forward to generating an unlimited number of unique and
unclonable ciphers without the maps having to be ‘designed by hand’. This makes available a solution to
encrypting data using one-time pads where an evolutionary computed cipher (and the key used to set the
initial condition as required) is used once and once only, after which it is discarded. In this context, a
known algorithm attack becomes a legacy of the past and conventional PRNGs such as those listed in [45]
may eventually become of historical interest alone.
The purpose of allowing an EC system such as Eureqa to iterate for what may be many hours on a
conventional CPU, is to evolve a cipher that provides a best approximation to the input noise. However,
for applications in cryptography, this may not be required as the output obtained after just a few iterations
may provide a ‘solution’ of the type specified by Equation (4) that generates suitable chaos, thereby
significantly reducing the evolutionary computational time required. Further, such maps can be
implemented on a complementary basis in which a data base of maps (constructed a priori) is accessed
so that different maps can be randomly selected and used to encrypt different blocks of data over a
randomized window of plaintext – multi-algorithmic cryptography.
Practical cryptography is based on passing known statistical tests available at [40], for example, which is
designed to ensure the pseudo-random property of a generator. Pseudo-random sequences are used
instead of truly random sequences in most cryptographic applications. This is because, a symmetric crypto
system can then focus on the application of a key to initiate an iterative formula, a formula that is known
to both Alice and Bob and preferably in the public domain. We have considered a method of designing
algorithms for generating pseudo random (chaotic) sequences using truly random strings to evolve an
iterator that is taken to be an approximation to these sequences. This approach pays no attention to the
algorithmic complexity of the iterator which is one of the main problems in the application of chaos to
cryptography. Neither does it consider the structural stability of the iterator or its algorithmic complexity.
However, it does provide a practical solution to the problem of developing a large database of PRNGs in
the application of personalizing encryption algorithms for strictly 1-to-1 communications or ‘1-to-Cloud’

Copyright © Society for Science and Education United Kingdom 37

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
(encrypted) data storage (e.g. [46], [47] and [48]). In this context, public knowledge of a specific
encryption algorithm becomes null and void.
By using EC systems such as Eureqa seeded with noise, it is possible to generate a nonlinear map with
automated control parameter settings. However, the one-step unpredictability does not guarantee that
the output sequence will be unpredictable when an adversary has access to a sufficiently long sequence.
In other words, the vast number of samples can, on a theoretical basis at least, lead to unpredictability.
With these provisions, EC has the potential for generating an unlimited number of ciphers which can be
personalized for users to secure their ‘Data on the Cloud’, for example. Evolutionary computed algorithms
can be published so that the approach conforms to the Kerckhoff-Shannon Principle as required, but in
the certain knowledge, that a new set of PRNGs can be developed by any individual interested in doing
so. In this context, the method might be viewed as a technical solution to the ‘democratization of the
cipher bureau’.
2.5.2 Encryption Methods
Maps of the type given in Table 2 output a floating-point vector 𝑥𝑥𝑘𝑘 to a precision that can be controlled
by the user, subject to the word length of the floating-point processor that is available. If the cipher is
taken to be a floating-point data field then for a ‘floated plaintext’ 𝑝𝑝𝑘𝑘 , the ciphertext will also be a floating-
point field, i.e. 𝑐𝑐𝑘𝑘 = (𝑥𝑥𝑘𝑘 + 𝑝𝑝𝑘𝑘 ). To decrypt the ciphertext, the recipient of the data must have access
to a processor that can provide the same floating-point accuracy. If this is not possible, the floating-point
field can be rounded to output a stream of integers so that for the 7-bit ASCII, the ciphertext is given by
Equation (1) where each component of this equation is a decimal integer, re-scaled as required so that
𝑥𝑥𝑘𝑘 ∈ [0,127], for example. Alternatively, this rounded cipher stream can be written as a binary stream
and the cipher computed using 𝒄𝒄 = 𝒙𝒙 ⊕ 𝒑𝒑. These encryption models through which a cipher is used to
actually encrypt the data are standard Vernam-type encryption algorithms adapted for different data
types. However, other encryption models can be used to encrypt the data as will now be discussed.
Instead of adding the plaintext to the cipher or implementing an XOR operation on the binary string
conversions, we can apply the convolution operation to a decimal integer or floating-point array. The
ciphertext is given by 𝒄𝒄 = 𝒙𝒙 ⊗ 𝒑𝒑 where ⊗ denotes the convolution sum and is an example of data diffusion
using convolutional encoding. Here, the arrays are taken to be vector arrays but convolutional coding can
be applied equally well for arrays or arbitrary dimension, most typically for encrypting images using the
two-dimensional convolution sum. Whatever the dimension of the process, decryption requires that the
ciphertext is de-convolved to recover the plaintext.
De-convolution is a well-known problem in digital signal and image processing and is often an ill-
conditioned problem because the process is unstable and requires to be regularized. However, for
cryptography, this problem can be solved by pre-conditioning the cipher with its own power spectrum. In
this case, the spectrum of the cipher denoted by 𝑋𝑋 and obtained by taking the Fourier transform of 𝒙𝒙
(which is a complex array with complex conjugate 𝑋𝑋 ∗ ) is replaced with the array 𝑋𝑋 ∗ /|𝑋𝑋|2 , |𝑋𝑋|2 > 0. Using
the convolution theorem, the convolutional coding process 𝒄𝒄 = 𝒙𝒙 ⊗ 𝒑𝒑 in data space becomes 𝐶𝐶 = 𝑋𝑋𝑋𝑋
in frequency space where 𝐶𝐶 and 𝑃𝑃 are the frequency spectra of 𝒄𝒄 and 𝒑𝒑, repectively. Thus, if we let
𝑋𝑋 ∗ 𝑃𝑃
𝐶𝐶 = (6)
|𝑋𝑋|2

URL: http://dx.doi.org/10.14738/tmlai.83.8219 38
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

then the spectrum of the plaintext is given by 𝑃𝑃 = 𝑋𝑋𝑋𝑋. In practice, this approach requires application of
the Fast Fourier Transform. Also, in practice, the condition that |𝑋𝑋|2 > 0 can be satisfied by letting |𝑋𝑋|2 =
1 if |𝑋𝑋|2 = 0 where the ′0′ is taken to be a floating point number whose value is below the floating point
accuracy available on a given processor. This method has been successfully applied for encrypted
information hiding (specifically, image Steganocryptography) – see [49], [50] and [51], for example. One
of the more useful properties that convolution-based encryption provides, is that it generates a ciphertext
that is data redundant. This facilitates a decrypt subject to a ciphertext with relatively significant data
error compared to an XOR process which requires bit error correction algorithms to be applied if the
ciphertext incurs errors due to transmission noise, for example.
Another method of encryption is to apply a phase-only process [52], [53]. In this case, the ciphertext is
obtain using the equation 𝑐𝑐𝑘𝑘 = 𝑝𝑝𝑘𝑘 𝑒𝑒 𝑖𝑖𝑥𝑥𝑘𝑘 , the decrypt then being given by 𝑝𝑝𝑘𝑘 = 𝑐𝑐𝑘𝑘 𝑒𝑒 −𝑖𝑖𝑥𝑥𝑘𝑘 . This is a process
that can be applied in any dimension and in data or frequency space. In data space, however, the
ciphertext is complex and therefore increases the data by a factor of 2. On the other hand, the plaintext
throughput can be doubled in size by using both real and imaginary components of a complex vector 𝒑𝒑.
It is of course possible to intercept the ciphertext, compute |𝑐𝑐𝑘𝑘 | and then attempt to retrieve the phase
from this data. However, the phase retrieval problem is severely ill-posed with no uniformly stable
solutions. Moreover, the practicality of implementing phase retrieval algorithms is dependent on the
dimension. It is well known that for the two-dimensional case, phase estimation algorithms have been
developed to provide approximate solutions. These solutions tend to rely on the data (whose amplitude
spectrum is known) being sparse. However, for the one-dimensional case, the phase retrieval problem is
ambiguousness, the determination of the phase within the extensive solution set being challenging and
only able to be considered under suitable a priori assumptions or additional information. This is a
consequence of the Fundamental Theorem of Algebra which states that every single-variable polynomial
with complex coefficients has at least one complex root. This theorem fails for polynomials of two
variables. The inability to factor polynomials of two variables is the reason why the two-dimensional
phase retrieval is possible and prevents the one-dimensional phase retrieval problem from being solved.
This is because the ability to factor polynomials generates ambiguities where multiple phases correspond
to the same data. Thus, any attack associated with attempting to solve the one-dimensional phase
retrieval problem will no doubt continue to remain a significant challenge for a cryptanalyst.
2.5.3 Natively Binary Chaos
The methodology discussed in Section 2.3 produces ciphers that are based on floating point iterators. In
some cases, the output of such iterators can be converted into binary streams either through necessity
(such as with the segmentation scheme illustrated in Figure 3) or through a preference to encrypt in binary
space (when developing real-time systems, for example). In the latter case, for example, any algorithm
that provides a positive only floating-point cipher 𝒙𝒙 ∈ [𝑥𝑥min , 𝑥𝑥max ], say, can be post-processed to
generate a cipher 𝒙𝒙 ∈ [0,1] using the equation
𝒙𝒙 − 𝑥𝑥min
𝒙𝒙: =
||𝒙𝒙 − 𝑥𝑥min ||∞
The binary string is then obtained by applying a round transformation which rounds each element of 𝒙𝒙 to
the nearest integer (in this case, 0 or 1). It is therefore natural to ask whether it is possible to design
algorithms that output binary strings directly without the need for conversion from decimal integer form,

Copyright © Society for Science and Education United Kingdom 39

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
such algorithms generating natively binary chaotic outputs. However, designing such algorithms is
relatively difficult compared to their floating-point counterparts which, as discussed in Section 2.3, are, in
the context of using EC, effectively unlimited. On the other hand, obtaining real world binary strings is
relatively easy and can be obtained through [41] and [42], for example. This leads to the idea of training
an ANN to output random binary streams for a few input (binary) strings (the keys). It also necessitates
the importance of obtaining tests that can accurately evaluate the randomness or otherwise of a binary
string which is considered in the following section.
A solution to designing an algorithm that can generate natively random binary strings is to apply the new
generation of Generative Adversarial Networks (GANs) first considered in 2014 [54]. A GAN is based on
using two neural networks which compete with each other in a zero-sum game. For a given training set,
the technique learns to generate new data but, crucially, with the same statistics as the training set. This
can of course include a binary string in which the distribution of a 0’s and 1’s is the same.

3 Binary String Analysis using Machine Learning

Just as an ANN can be trained to generate random binary strings, so it can be used to evaluated the
randomness (or otherwise) of the string. In principle this could be done using a deep learning strategy as
discussed in Section 1.3. However, of all the tests on randomness, the information entropy is a metric
that is especially sensitive to the random nature of data in general. It is fundamental to quantifying the
meaning of information and hence, the lack of it through, for example, the randomization of a plaintext
into a ciphertext with an equal probability of the states. This is why, ciphers that are uniformly distributed
are referred to as maximum entropy ciphers. In this section we consider a binary entropy test (a
modification thereof) that produces a statistically significant difference between a non-random and a
genuinely random binary string and further, can assess the degree to which they are one or the other.
A long standing problem in regard to the analysis of a finite binary string or bit stream (of compact
support) is how to tell whether the string is representative of non-random (intelligible) information
(involving some form of periodicity, for example) or whether it is the product of an entirely random
process or whether it is something in between the two. This problem has applications that include
cryptanalysis, quantitative finance, machine learning, artificial intelligence and other forms of signal
processing involving the general problem of how to distinguish real noise from information that is
embedded in, or is entirely lost, in noise. In cryptanalysis, given that digital communications is based on
transmitting bit streams, it is routine to analyze the streams in an attempt to identify a signature which
reflects whether the information is encrypted (or otherwise) and if so, the strength of the encrypted
signature that is present in the binary data. We now consider the basis for solving this problem using an
entropy-based analysis that gives a statistically significant difference between non-random and genuinely
random binary strings. The statistical significance is compounded in a marked difference in the distribution
of the output signals from which some basic statistical metrics can be computed and used to train an ANN,
thereby developing a machine learning strategy to analyze the randomness of a binary stream as it
changes in time.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 40
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

3.1 Information Entropy

In Leo Szilard's 1922 doctoral dissertation and companion landmark paper, he showed how a thought
experiment concerned with the second law of thermodynamics called the Maxwell Paradox, could be
solved by introducing the concept of information entropy which provides a ‘balance' to the paradox
associated with a perceived decrease in the Boltzmann entropy [55]. Szilard's original concept on
information entropy has become the basis of information theory, showing that there is an increase of
𝑘𝑘𝐵𝐵 log 2 2 units of entropy in any measurement where 𝑘𝑘𝐵𝐵 is the Boltzmann constant. This concept was
independently re-invented by Claude Shannon in 1949 [56] (to whom credit is usually but erroneously
given) and Andre Kolmogorov and Yakov Sinai, who developed a modified form later on in 1959 [57], [58].
In developing a solution to a paradox in thermodynamics, Szilard introduced an idea that is arguably the
single most important icon of the information revolution and the digital age. This is because information
entropy provides the key for estimating the (average) minimum number of bits needed to encode a string
of symbols, based on the frequency of those symbols.
In a more general context, information is a measure of order, a universal measure applicable to any
structure or any system. It quantifies the instructions that are needed to produce a certain organization.
We compute the information inherent in any given arrangement from the number of choices we must
make to arrive at that particular arrangement among all possible arrangements. Intuitively, the more
arrangements that are possible, the more information that is required to achieve a particular
arrangement. There are several ways in which one can quantify information but an especially convenient
one is in terms of binary choices which is quantified through the binary information entropy.
Consider a digital signal 𝑠𝑠𝑘𝑘 , 𝑘𝑘 = 1, 2 … . , 𝐾𝐾 composed of 𝐾𝐾 (real) values. Let the discrete probability
distribution (constructed from 𝑁𝑁 bins) that a specific value 𝑠𝑠𝑘𝑘 occurs in the signal be 𝑃𝑃𝑛𝑛 , 𝑛𝑛 = 1, 2, . . . , 𝑁𝑁.
The information associated with an outcome (i.e. a specific value of 𝑠𝑠𝑘𝑘 with a bin) is then − log𝑃𝑃𝑛𝑛 which
is a measure of the information required to specify 𝑠𝑠𝑘𝑘 in terms of it being a member of a bin where 𝑃𝑃𝑛𝑛 is
the distribution of bins. In this context, the (Shannon) Information Entropy (usually denoted by 𝑆𝑆), is a
measure of the mean (the ‘expected value') of the information measure − log𝑃𝑃𝑛𝑛 and is given by the
dimensionless quantity
𝑁𝑁

𝑆𝑆 = − � 𝑃𝑃𝑛𝑛 log𝑃𝑃𝑛𝑛
𝑛𝑛=1

The higher the entropy of a signal becomes the greater its ambiguity, and, in this context, information
entropy is a measure of the unpredictability or randomness of any message contained in the signal. This
is typically determined by the noise that distorts the information contained in a signal. In general, the
information entropy associated with the transmission of information in a signal tends to increase with
time. This is due to the increase in noise with time that distorts any signal as it propagates in time, the
sources of this noise being multi-faceted but tending to a Gaussian distribution as a consequence of the
Central Limit Theorem.
Instead of considering a digital signal composed of decimal integer or floating-point elements, let us now
consider the signal to be a binary string whose elements can take on only two values, namely 0 or 1, which
are mutually exclusive. In this case, the signal will have a binary distribution 𝑃𝑃𝑛𝑛 , 𝑛𝑛 = 1,2 consisting of just
two bins and the Binary Information Entropy (BIE) function, denoted by 𝐻𝐻, becomes

Copyright © Society for Science and Education United Kingdom 41

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
𝐻𝐻 = − ∑2𝑛𝑛=1 𝑃𝑃𝑛𝑛 log 2 𝑃𝑃𝑛𝑛 = 𝑝𝑝log 2 𝑝𝑝 − (1 − 𝑝𝑝)log 2 (1 − 𝑝𝑝) Bits (7)
where, if we let 𝑝𝑝 denote the probability of 1 occurring in the binary string, then the probability of
obtaining a 0 in the same string is 1 − 𝑝𝑝. Similarly, if 𝑝𝑝 is taken to denote the probability of 0 occurring
in the string, then the probability of obtaining 1 is 1 − 𝑝𝑝. In either case 0log 2 0 ≡ 0.

3.2 Evaluating the Order/Disorder of a Binary String

Given a binary string of a finite length, our problem is to evaluate whether the string is a binary
representation of genuine noise or whether it contains intelligible information in terms of it having some
degree of determinism. The determinism of such a string could include any natural language that has
evolved through use, application and repetition without conscious planning but binary coded in a planned
and premeditated way, e.g. the ASCII. The purpose is therefore to establish a method by which a finite
binary string of arbitrary length can be compared against another in terms of the relative order and/or
disorder of all of its bits. Applying a basic binary entropy test is not sufficient. This is primarily because of
its failure to differentiate between binary strings that include periodicity and are therefore not random.
For perfectly ordered binary strings whose elements are all equal to 1 or all equal 0 and when 𝑝𝑝 = 0 or 𝑝𝑝 =
1, 𝐻𝐻(𝑝𝑝) = 0. On the other hand, when 𝑝𝑝 = 0.5, 𝐻𝐻(𝑝𝑝) = 1 reflecting maximum irregularity. However, for
a string such as 01010101 which has a repeating pattern of 01, 𝑝𝑝 = 0.5, 𝐻𝐻(𝑝𝑝) = 1 so that in this case,
the value of 𝐻𝐻(𝑝𝑝) appears to represent a random binary string when in reality, the sting is periodic and
therefore not random at all. Thus, what is required is a binary information entropy-based metric that
differentiates more fully than is possible with Equation (7) alone. This issue is further quantified in Table
2, which gives exemplars in regard the order and disorder of some 8- bit binary strings.
Table 3. Some exemplars of order and disorder associated with some short binary strings (from [68]).

Example Binary String Description Reason

11111111 Perfectly ordered All 1’s
00000000 Perfectly ordered All 0’s
01010110 Mostly ordered Mostly 01’s
01010101 Regular, not disordered Repeating 01’s
11001100 Regular, not disordered Repeating 1100’s
01011010 Mostly ordered 0101 then 1010
01101011 Somewhat disordered No apparent pattern
10110101 Somewhat disordered No apparent pattern

There have been a number of algorithms designed to compute various entropy-based metrics for the
determination or measurement on the randomness, regularity, irregularity, order, disorder and entropy
for binary and other strings [59], [60]. These include measures that aim to characterize randomness and
disorder through the entropy of finite strings such the Approximate Entropy [61], [62], Sample Entropy
[63] and Fuzzy Entropy [64]. Such metrics are based on algorithms that can be classified as follows: (i)
moving window methods that examine sub-strings of the original; (ii) algorithms which generate a metric
based on the entire string length. Applications include basic randomness tests [65], [66] and cryptanalysis
[67]. This includes the development of a BiEntropy measure which is based upon a weighted average of
the Shannon entropy for all but the last binary derivatives [68].
These metrics are the product of many studies that have been and continue to be undertaken to develop
suitable tests and measures based on the computation of a single metric. While desirable

URL: http://dx.doi.org/10.14738/tmlai.83.8219 42
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

computationally, focusing on the use of a single metric for this purpose is restrictive and may be
statistically insignificant because of its self-selecting data predication. For this reason, we consider a
complementary approach to the problem which is based on the application of the Kullback-Leibler
Divergence for a stream of data that yields a statistically significant result as opposed to a single metric.
This provides the foundations for the application of a machine learning approach.

3.3 Distribution Analysis using the Kullback-Leibler Divergence

Since randomness is a relative concept, a relative metric is considered which provides a measure of how
a binary string compares in some way with a string that is known to be the product of a genuinely random
process. Further, this comparison needs to be undertaken on a statistical basis, measuring how one
probability distribution associated with the binary string compares to a reference probability distribution
in terms of its information content. For the reason, we consider the Kullback-Leibler Divergence or
Relative (Binary) Entropy function given by
2
𝑄𝑄𝑛𝑛
𝑅𝑅 = − � 𝑃𝑃𝑛𝑛 log 2 � �
𝑃𝑃𝑛𝑛
𝑛𝑛=1

where 𝑃𝑃𝑛𝑛 is the binary histogram of binary string 𝑓𝑓𝑘𝑘 and 𝑄𝑄𝑛𝑛 is the binary histogram of some reference
binary string 𝑔𝑔𝑘𝑘 , both binary strings being finite and of the same length.
Suppose the string 𝑓𝑓𝑘𝑘 is mostly ordered or regular and not disordered (e.g. a binary string representation
of some text from a natural language) and 𝑔𝑔𝑘𝑘 is a random string. We require the metric 𝑅𝑅 to be
significantly different in terms of its numerical value to the case when both 𝑓𝑓𝑘𝑘 and 𝑔𝑔𝑘𝑘 are random binary
strings. Ideally, what is required is to establish a threshold for the value of 𝑅𝑅 below which 𝑓𝑓𝑘𝑘 can be
classified as ordered say, and above which, 𝑓𝑓𝑘𝑘 can be classified as random. This is an example of a binary
decision-making process that may not be statistically significant and may not necessarily be able to
‘monitor’ a transition from 𝑓𝑓𝑘𝑘 being random to non-random. Thus, suppose we consider the relative
entropy digital signal given by
𝑄𝑄
𝑅𝑅𝑚𝑚 = − ∑2𝑛𝑛=1 𝑃𝑃𝑛𝑛𝑛𝑛 log 2 � 𝑃𝑃𝑛𝑛𝑛𝑛� , 𝑚𝑚 = 1, 2, … , 𝑀𝑀 (8)
𝑛𝑛𝑛𝑛

where 𝑃𝑃𝑛𝑛𝑛𝑛 is the 𝑚𝑚th binary histogram of the input binary string with 𝑃𝑃𝑛𝑛1 = 𝑃𝑃𝑛𝑛2 = ⋯ = 𝑃𝑃𝑛𝑛𝑛𝑛 and 𝑄𝑄𝑛𝑛𝑛𝑛 is
the 𝑚𝑚th binary histogram of an 𝑚𝑚th random binary string obtained using a PRNG or is obtained from a
real-world random binary string source. To illustrate the characteristics of this relative entropy signal, we
consider the following cases (which are referred to as such in regard to presenting the results that follow):

Case (i): 𝑃𝑃𝑛𝑛𝑛𝑛 is the 𝑚𝑚th binary histogram of a non-random string;

Case (ii): 𝑃𝑃𝑛𝑛𝑛𝑛 is the 𝑚𝑚th binary histogram of a random string.

In the results that follow, the non-random string is obtained by generating the binary representation of
the text associated with the abstract of this paper, achieved using an ASCII text to binary converter [69]
(with the delimiter set to none). A sequence of random binary arrays are generated using the MATLAB
uniform distributed random number generator function rand (which returns floating point numbers in the
interval [0,1]) and applying a round transformation (to output an array consisting of 0's and 1's), each
array having the same length and each array being independent of the another.

Copyright © Society for Science and Education United Kingdom 43

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
Figure 5 shows example signals of the relative entropy given by Equation (8) for Case (i) and Case (ii) above
with 𝑀𝑀 = 8000 and 100-bin histograms. It is then clear that:

• For Case (i), when 𝑓𝑓𝑘𝑘 , is a non-random string, 𝑅𝑅𝑚𝑚 and has a Gaussian-type distribution.
• For Case (ii), when both strings are random, 𝑅𝑅𝑚𝑚 also has a Gaussian-type distribution.
In both cases, application of the Jargue-Bera test for normality shows that the null hypothesis must be
rejected, i.e. the series 𝑅𝑅𝑚𝑚 does not actually conform to a normal distribution even though it appears to
have the characteristics of one (as given in Figure 5). Nevertheless, Figure 5 demonstrates the ability for
Equation (8) to provide a statistically significant measure in regard to evaluating the randomness or
otherwise of a finite binary string. The difference in the mean 𝑅𝑅� of the relative entropy signal for the two
cases is at least two orders of magnitude and can therefore easily be used to differentiate between a
random and non-random binary string. But this is just one of many statistical metrics that may be
computed from 𝑅𝑅𝑚𝑚 by computing its J-bin histogram 𝐻𝐻𝑗𝑗 , 𝑗𝑗 = 1, 2, . . . , 𝐽𝐽 and evaluating the 𝑟𝑟 th central
moment (the moments about the mean 𝑅𝑅� ) given by
𝐽𝐽

µ𝑟𝑟 = �(𝑗𝑗 − 𝑅𝑅� )𝑟𝑟 𝐻𝐻𝑗𝑗

𝑗𝑗=1

providing the variance (𝑟𝑟 = 2), the skewness (𝑟𝑟 = 3) and Kurtosis (𝑟𝑟 = 4), for example, coupled with
metrics such as the median and the mode given by ||𝐻𝐻𝑗𝑗 ||∞ .

Given the demarcation between a random and non-random binary string using Equation (8), the potential
exists to compute further statistical metrics and other parameters based on an analysis of the signatures
given in Figure 5. These may include the higher statistical moments and spectral properties of the digital
signal 𝑅𝑅𝑚𝑚 , for example, designed to develop a feature vector whose purpose is to provide a multi-class
classification using an ANN starting with a simple four component feature vector consisting of the Mean,
Standard Deviation, Median and Mode.
Since all data can be expressed as a binary string, irrespective of the code used to do so (i.e. ASCII or
otherwise), the approach compounded in utilizing Equation (8) provides a relatively generic method of
differentiating between random and non-random binary strings. Unlike other approaches to solving this
problem, this approach is based on generating a signal computed using Equation (8) and characterizing
the distribution of the signal rather than computing a single metric for a binary string. This allows some
common statistical metrics to be used to classify changes to the distribution of the signal 𝑅𝑅𝑚𝑚 and how
these changes can form the basis for a machine learning approach using complementary statistical
parameters for the signal, coupled with metrics associated with the spectrum of the signal and other
transformations. The purpose of this is to provide an analysis of a binary stream that can specify whether
it is truly random, intelligible (i.e. the product of some communicable information, for example) or
partially random in some way. This provides the ability to continuously monitor binary streams to test
for plaintext (intelligible) or encrypted (non-intelligible and random) data.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 44
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

Figure 5. Plots of the 𝑅𝑅𝑚𝑚 (left) given by Equation (8) for 𝑀𝑀 = 8000 and the associated 100-bin histograms
(right) for Case (i) (above) and Case (ii) (below), respectively.

4 Case Study: Document Authenticity using Unclonable Ciphers

Technologies to make high value documents authentic and impossible or at least difficult to counterfeit
have been pioneered for many years. This includes the use of foil holograms, micro-printing, raised print,
ultra-violet features, overt and covert watermarks as well as the more recent use of polymers instead of
paper for bank notes, for example. With these developments have come a range of specialist readers for
checking the authenticity of bank notes at a Point of Sale (PoS) and passports at airports, for example. In
this case study, we consider the use of unclonable ciphers based on the application of EC for
authenticating documents, specifically with a smartphone. This is because smartphones are now very
common with some 3.5B users world-wide (45.12% of the current world population; 2.5B in 2016) and
because it is trivial to download a specialist app including those based on the methods discussed here.
Consequently, there is no added expense for authentication at a PoS; it is simple to introduce new apps
for upgrades and obviates the need to buy specialist readers. There are two potential ways of
authenticating a document with a smartphone: (i) using the radio frequency antenna; (ii) using the
camera. In this case study we discuss new technologies for both.

4.1 Radio Frequency Identification

Suppose that an Integrated Circuit (IC) coupled to an antenna can be designed that is thin and flexible
enough to be embedded into a document and includes enough read only memory to store an unclonable
ciphertext that is unique to the IC. In the near field, the antenna of a smartphone can then be used to
both activate and read the ciphertext, thereby authenticating the document on-line. The key to this form
of RFID (Radio Frequency Identification) [70] is to ensure that the encrypted information maintained on
the IC, is unique and unclonable. This requires the generation of numerous chaos-based functions that
can be obtained using EC.
FLEX NFCTM stamp antennas [71] are a new generation of ultra-thin (up to 10 µm compared to standard
IC’s of 75µm or 120µm) low cost and flexible integrated circuits made of amorphous silicon deposited on
a polyamide substrate. This extends electronics beyond the realm of conventional mono-crystalline

Copyright © Society for Science and Education United Kingdom 45

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
silicon-based IC’s. Flexible IC’s can be applied to various packaging and security documents, for example,
and can be embedded in polymer-based bank notes, for example, using a similar approach to embedding
foil holograms. They include a near field communication antenna with typical operational frequencies of
13-14 MHz, a self-resonance frequency of 100 MHz, an impedance of 50-80 Ohms and a reading distance
of up to 40 mm.
Authenticity is provided by using physically unclonable maps, a method to create ciphertexts that are
inextricably linked to the device. The ciphertexts are unique for a given device, are impossible to read,
suspect or eavesdrop on, and any interference by the device destroys them. In addition to EC, this solution
can be based on chaos theory which enables the extensions of the dispersion of physical parameters of
electronic circuits. The solution's innovation lies in the exploitation of time instability which guarantees
that the ciphers and keys are random and repeatable only for a specific device [72]. The solution is
implemented as a hardware based non-programmable coprocessor which forms part of the flexible IC and
is characterised by a maximum entropy encrypted module with low physical complexity and low power
consumption coupled to Internet of Things (IoT).
Figure 6 provides a schematic of the approach for the authentication and identification of a bank note.
This is based on correlating a unique ciphertext maintained of a flexible IC (embedded in the note) and
the serial number of the bank note. The ciphertext is an encryption of the serial number on the bank
note, for example, but other data relating to the note can be included as required. The cipher used for
encryption can be an EC generated cipher which is unique to the serial number and unclonable. The
ciphertext is read by the antenna of a smartphone (in the near field), transmitted by the phone using the
IoT to a secure relational data base where the encrypted data is identified. If identified successfully, the
unique cipher is recovered and used to decrypt the data recovering the serial number of the banknote.
This result is then transmitted back to the phone quantifying the authenticity or otherwise of the bank
note while displaying the serial number as required, and, as illustrated in Figure 7.
In the future development and deployment of this technology, the application of EC based nonlinear
functions for generating the many hundreds of millions of unique unclonable ciphers required to embed
a unique ciphertext is clear. For any evolved function given in Equation (4), the initial condition can be set
to the serial number of the bank note, for example. Each function is ‘encoded’ in the flexible IC during a
production run. Decryption and validation (or otherwise) is accomplished automatically by searching for
the unique cyphertext held in a relational data base and its correlation with the serial number upon
reception from a smartphone over the IoT.

Figure 6. Schematic illustration of the application to a bank note of a flexible IC with a unique ID number
(encryption cipher/key) maintained on a non-programmable coprocessor with a ROM.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 46
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

Figure 7. Example of the IoT return for a banknote with an embedded flexible IC using a smartphone to read
and transmit the ciphertext and provide the user with a decision on the authenticity of the banknote, i.e.
GENUINE (left) or counterfeit and NOT AUTHORISED (right).

4.2 Optical ID
In addition to using online authentication technologies associated with unique RFID’s on a smartphone, it
is also possible to consider active ‘optical solutions’ using the digital camera of a smartphone to obtain
digital images in ambient light and/or flash light conditions. The retro-reflector method described here is
based on the latter case. Compared to RFID, optical ID provides the opportunity to develop a range of
optical ciphers and associated identifiers, but this comes at the cost of having to process and transfer
digital images using the IoT which requires higher data throughputs than those associated with a RFID.
On the other hand, relatively simple print-only and/or additive optical elements and complexes can be
included that are simple and cost effective as discussed in this section.
4.2.1 Printed Binary Texture Coding
We consider an approach which is based on the introduction of binary texture codes to a document.
Texture codes are based on applying convolutional coding to an image plaintext based on Equation (6) to
produce a ciphertext image [51]. The method is relatively robust to distortions of the ciphertext accept
in regard to three issues, size, rotation and cropping. In other words, if the ciphertext is printed on a
document and then scanned, the scanned image must be cropped correctly, re-oriented (as necessary)
and then re-sampled back to the original size of the plaintext image. However, auto-orientation and auto-
cropping are routine features of most scanning and remote imaging systems as is the re-sampling of a
digital image. Moreover, if the plaintext image is binary, the texture code is data redundant and can
therefore can be binarized. Correlation of this binary texture code with the original cipher then recovers
the binary plaintext with background noise but with a high signal-to-noise ratio thereby allowing the
plaintext to be identified.
Figure 8 shows an example of binary texture coding for a Euro banknote. In this case, the serial number
has been used as the binary plaintext, and the binary texture code printed in ultraviolet ink over the entire
surface of the banknote at 150 dpi. Under ultraviolet light, an image of the texture code is captured and
upon, auto-cropping, auto-correcting (for orientation, as required) and re-sampling the image, the result
is decrypted to recover the serial number. In this case, an EC cipher can be used and iterated, according

Copyright © Society for Science and Education United Kingdom 47

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
to Equation (4), with double precision and a key equal to the serial number, i.e. 𝑥𝑥0 = 0.00000004795.
The reason for printing the binary texture code using ultraviolet ink is to make the code invisible under
visible light. This approach is of course similar to using QR codes, but it provides the basis for incorporating
greater information content subject to compatibility with the physical size of the window that is used to
print the texture code and the dpi coupled with the resolution of the image capture device.

Figure 8. Example of binary texture coding applied to authenticate a bank note. Original note (left), binary
texture code (center) and decrypt to recover the serial number (right).

4.2.2 Laser Speckle Analysis

All surfaces have a degree of roughness over different scales and even a highly polished surface has
surface roughness at the micro- or nano-scale. The term roughness used here refers to the random
undulations on the scale over which the roughness is considered to occur. On the scale of the wavelength
of light (400 to 700 nanometres or nm) the roughness of a surface is quite pronounced. This creates
internal surface scattering effects. Even with highly polished surfaces such as those required for parabolic
mirrors and other optical components, surface irregularities exist on the scale of a wavelength of light.
These irregularities are typically measured in terms of the Root Mean Square (RMS) irregularities and
surface imperfections (Scratch and Dig) that are typically of the order of tens of nanometres. In the
Hubble Space telescope, for example, the primary mirror was polished to give a 0.014-wave RMS wave-
front error at 632.8 nm [73]. In addition to maintaining an accurate geometry, the smoothness associated
with the surface of optical components on the scale of a light wave is vitally important in order to generate
high optical resolutions. The RMS irregularity of a surface determines the characteristics of the Point
Spread Function (PSF), in particular the effective bandwidth of the Optical Transfer Function (OTF) - the
Fourier transform of the PSF [15]. As the RMS irregularity increases the bandwidth decreases due to
surface scattering effects distorting primarily the high frequency band of the OTF.
The scattering effects from a random surface of incoherent light are not phase-sensitive but with coherent
light, the phase associated with the scattering processes are observable and measurable and produce a
pattern known as a speckle pattern. This is because, with coherent light, the wave-fronts interfere
mutually. Speckle patterns are generated by the diffuse reflections of monochromatic light (generated
by a laser) which involve multiple scattering events when the scattered optical field can be assumed to be
similar to the process of diffusion due to an ensemble of particles undergoing random walks. The diffuse
surface reflection of coherent light is generated by any rough surface (i.e. rough on the scale of the
wavelength) or media with a complex of scattering objects (with a scale length similar to the wavelength)
which have a different permittivity and/or conductivity to the media in which they are suspended. This
includes the speckle patterns generated by the surface of paper and other materials used to produce
security documents.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 48
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

The key to using Laser Speckle Analysis for document authentication is to appreciate that the speckle
pattern is a signature of the surface topology (on the scale of the wavelength) associated with a chosen
surface patch. Moreover, if the physical parameters associated with the incident coherent radiation can
be reproduced accurately enough (including wavelength, angle of incidence of the laser beam to the
surface, the beam profile and the angle, relative to the surface at which the speckle pattern is recorded)
and the surface patch is not deformed over time, then the speckle pattern provides a unique signature of
the patch on the document, thereby providing a method of authentication. Further, this method of
authentication is passive given that the surface structure of the patch is a natural feature of the material
from which the document is composed. All that is required is to decide upon the position and the extent
of the surface patch whose speckle pattern signature is desired and then record the pattern produced for
a specific optical configuration.
Figure 9 shows a simulated speckle pattern and the characteristic distribution of the pattern plotted on a
logarithmic scale. This simulation is obtained using Equation (6) without power spectrum normalisation,
i.e. we use 𝐶𝐶 = 𝑋𝑋 ∗ 𝑃𝑃, where the Plaintext spectrum 𝑃𝑃 is an OTF given by a two-dimensional rectangular
function and 𝑋𝑋 is the spectrum of a zero-mean Gaussian distributed random field. The speckle pattern is
then given by the absolute square of the inverse Fourier transform of the cipher 𝐶𝐶 which models the
scattering cross-section (the intensity of scattered light).
The distribution of the scattering cross-section (the observed laser intensity, 𝐼𝐼 ≥ 0) associated with a
speckle pattern is well known and given by the (negative) exponential distribution
𝑃𝑃(𝐼𝐼) = 𝜆𝜆 exp(−𝜆𝜆𝜆𝜆)
where 𝜆𝜆 is the ‘rate’. Thus, under a logarithmic transformation,
ln 𝑃𝑃(𝐼𝐼) = ln𝜆𝜆 −𝜆𝜆𝜆𝜆,
a relationship that is well demonstrated in the logarithmic scaled histogram given in Figure 9.
By binarizing the speckle pattern, it is possible to generate a code similar to a QR code but one that is
produced naturally by the interaction of coherent light with a surface. Such a speckle code can be viewed
as a cipher, a cipher that is a unique signature associated with a known surface patch and optical
configuration. A speckle pattern can be considered to be another natural noise source in addition to those
discussed in Section 2.3.1, for example, and, could be used as an input to the EC systems and approach
discussed in Section 2.3. An optical cipher of the type simulated in Figure 9 could be used to encrypt the
serial number of a bank note to produce a binary texture code following the approach discussed in Section
4.2.1, for example.

Copyright © Society for Science and Education United Kingdom 49

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60

Figure 9. Simulation of a speckle pattern (top-left) and the corresponding grey-level 100-bin histogram (top-
right) plotted on a logarithmic scale. A binarization of the speckle pattern designed to produce a matrix of 0’s
and 1’s of near equal population densities is shown in the lower left-hand image. The corresponding 2-bin
histogram of this image is given in the lower right-hand side bar graph.

While the methods discussed in this section and that of Section 4.2.1 can be implemented in practice,
they both require ‘laboratory conditions’ to be realised. They require the use of UV or laser light to
generate and recover the authentication patterns, patterns that need to be recovered, processed and
analysed under controlled conditions with specialist equipment and minimal error in terms of an optical
recording configuration. Although the analysis of such patterns for authentication could be considered
using Deep Learning, for example, it is clearly not currently possible to implement such an approach in
the field using smartphones which are not equipped with UV and laser light sources. Thus, in the following
section, another approach is considered that is compatible with the optical facilities provided by a
smartphone and can be operated in the field.
4.2.3 Micro-Retro Reflector Texture Coding
Although binary texture codes can be generated without UV printing, a UV print provides a way of hiding
the existence of the code under visible ambient lighting conditions. In this way the cipher remains hidden
in the optical spectrum. Another way to do this (i.e. introduce a cipher that is invisible under ambient
lighting conditions) is to use a random distribution of micro retro-reflectors which is set into the surface
of the document providing a random pattern that is unique to that document. Under ambient lighting
conditions, the retro-reflector complex will not be observed visually and will not be detected in a captured
digital image of a smartphone. However, with a halogen flash that is typically coupled with most modern
smartphone cameras, the retro-reflectors will reflect the light and their distribution (or rather the
combined effect of that distribution) on the document captured in the digital image that is retained.
Retro-reflectors are devices that reflect light back to its source with minimal scattering [74]. There are
many types of such devices including corner reflectors, prism-based reflectors and phase-conjugate
mirrors but the most common type makes use of micro retro-reflectors which are each composed of small
glass beads or micro-spheres. The beads vary in size from 0.1-1 mm and can be added to paint and
lacquers which then cause the surface of the material to which the paint/lacquer has been applied to
glow. Highway signs and painted stripes on roadways, for example, have been coated with micro-
spherical retro-reflectors for many years. Various reflective sheets are used to coat signs, some of which
employ layers of glass beads, while others use sheets embedded with micro-plastic corner prisms or
micro-prisms.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 50
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

Figure 10 shows an example of a bank note which has been coated with Glowtec fine grade retro-reflective
powder (a pigment for use in spirit and water-based mediums) [75] over a rectangular window within
central component of the main graphic. The powder is composed of retro-reflective glass microspheres
and has been introduced by adding the powder to a clear nail varnish and then applying the varnish to the
surface of the note with the (nail varnish) brush. The size of the microbeads is 30-40 microns each with a
refractive index of 1.93 and its application to the surface using a clear lacquer produces a naturally random
distribution of microbeads which are unique for any particular note. Under ambient lighting conditions,
the existence and distribution of the micro-spheres is not visible and the captured image of the bank-note
on a smartphone is ‘normal’. However, application of a halogen flash reveals the rectangular region over
which the powder doped lacquer has been applied. Thus, by cropping this region of the image, a unique
optically generated cipher is obtained which is a combination of the physical micro-bead distribution
subject to the resolution at which the image is captured.

Figure 10. Smartphone images of a banknote with reflective powder applied under ambient lighting
conditions (left) and with a LED halogen flash (right). The reflective power, consisting of a randomized complex
of micro retro-reflectors, has been applied by doping a clear nail varnish with the powder and ‘painting’ it onto a
rectangular patch that covers the central component of the notes principal graphic as shown on the example on
the right-hand-side.

Diffusing a grey level version of this image using Equation (6) with a critical feature of the banknote such
as a binary image of the serial number, a ciphertext can be generated and stored in a data base. This
ciphertext is then a uniquely encrypted record of the banknotes serial number coupled with the unique
optical cipher on the same bank note which remains invisible under ambient lighting conditions. Figure
11 shows example images obtained by applying this process, namely, the cropped optical cipher, the
ciphertext for the serial number and a decrypt obtained by re-imaging the cipher. As discussed in Section
4.2.1, the method of encryption by convolutional coding means that the image of the cipher is relatively
tolerant to distortion other than orientation, cropping and compatibility with the original digital image
size (and the cipher).

(a) (b) (c) (d)

Figure 11. Grey level image of the cropped optical cipher (a), the ciphertext (b) obtained by diffusing the
cropped optical cipher with a binary image of the serial number (c) and the decrypt (d). Note that the binary
image of the serial number is deliberately corrupted with uniformly distributed noise in order to enhance the
uniformity in image diffusivity of the ciphertext shown (a).

Copyright © Society for Science and Education United Kingdom 51

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
The method is also tolerant to changes in the angle at which the image is taken relative to the plane of
the banknote. This because the retro-reflectors reflect light back to the source irrespective of the angle
of incidence subject to distortions obtained at angles greater than the order of 30 degrees or more.
A typical PoS application of this technique in order to authenticate a bank note is as follows:
• An image of the bank note is taken on a smartphone with the LED flash.
• If the image contains a bright field area (as shown in Figure 10, for example) then the note is taken
to include an optical ID; this feature alone being a form of low-level authentication.
• The optical ID is auto-cropped and the cropped image submitted on-line where it is located in a
relational database of such images. Note that the crop can include features that lie inside or
outside the region in which the optical cipher is placed as long as there is consistency between
the crop obtained and that is used to construct the database, the basic principle being to
introduce unique randomness into the captured image.
• The cropped image is resampled and decrypted (i.e. a grey level image of the serial number)
returned back to the smartphone coupled with image processing as required to enhance the visual
quality of the decrypt that is displayed. This is required because the binary plaintext is deliberately
corrupted with uniform noise whose purpose of this is to enhance the uniformity of image
diffusivity in the ciphertext.
• The decrypt can then be compared with the serial number of the banknote in order to
authenticate it.
In this case, the cipher is the image of the retro-reflective powder patch and does not, in principle, require
further encrypting. However, in order to strengthen the ciphertext further, it is possible to double encrypt
based on extending Equation (6) to the form
𝑌𝑌 ∗ 𝑋𝑋 ∗ 𝑋𝑋 ∗
𝐶𝐶 = |𝑌𝑌|2 |𝑋𝑋|2
𝑃𝑃 or 𝐶𝐶 = 𝑒𝑒 𝑖𝑖𝑖𝑖 |𝑋𝑋|2
𝑃𝑃

where 𝑌𝑌 is the spectrum of an evolved cipher using EC whose key could also be the serial number of a
bank note.
In this application, ANNs will be of value in regard to the automatic recognition of the optical cipher and
digital image post-processing that may be required in order to provide a ciphertext that is suitable for
decryption. The purpose of this is to allow the recovery of the optical ID to be undertaken in non-ideal
conditions associated with the way in which users operate a smartphone subject to critical tolerances.
This includes optical clarity, depth and resolution coupled with excess distances from the banknote at
which the image is taken, for example, and other issues associated with inevitable mismanagement in the
use of an app which is critically dependent on the quality of image capture.

5 Summary, Conclusions and Future Directions

The purpose of this paper has been to provide a broad overview of how AI is starting to be applied in the
development of encryption algorithms and cryptanalysis. We have considered the role that Evolutionary
Computing can play in providing personal ciphers in the form of unclonable chaotic maps and how, in a
complementary way, ANNs can be trained to simulate chaos. In both cases, the training data is based on
access to real-world noise sources, which include service providers such as Random.org. This approach
yields the potential for a ‘democratization of the Cipher bureau’ where personalized ciphers can be

URL: http://dx.doi.org/10.14738/tmlai.83.8219 52
 Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

created with relative ease without having to resort to ‘hand-crafted’ ciphers, many of which are now
vulnerable to attack by the implementation of Shor’s algorithm [76] on a quantum computer. This is
because encryption algorithms that are predicated on the product of large prime numbers can be broken
using Shor’s algorithm, but until the development of quantum computing, this algorithm was not time
computational feasible. In this respect, the approach considered in the paper falls in to the category of
Post-quantum Cryptography and the generation of software for prototyping quantum resistant
cryptography [77].

5.1 Conclusion
In a broad context, and, in regard to the use of AI for cryptography, there is an interesting analogy that
can be made with the work of Station-X at Bletchley Park during the second world war. In this case, the
problem was to break the key settings of the Enigma cipher used by the German Armed forces for
communications. However, the design of the cipher was known, as was the language being encrypted and
many of the expected and repetitive words, phrases and statements (passive Cribs). Moreover, the
intercepted Morse coded traffic was known to be encrypted because it was scrambled. Had the traffic
been disguised to appear routine through encrypted information hiding (not technically possible at the
time), the Morse code have been reconfigured (quite possible at the time), the language been translated
to a non-Indo-European language (possible at the time but requiring multi-lingual service providers),
repetitive words a phases negated (possible) and the Enigma machine itself changed on a regular basis
(not technically possible at the time), then history might have been very different (albeit not necessarily
tolerable). It is the last component of this analogy that is the basis for using AI to design ciphers that can
be changed regularly or even changed for each communication to give an algorithmic one-time pad.
Because this is not yet practicable for all individuals (many of whom may lack the technical IT skills), in the
interim, there is ample potential for business opportunities based on establishing on-line service
providers to distribute personalized ciphers for 1-to-Cloud applications, 1-to-1 encrypted communication
and 1-to-many applications depending on the available distribution infrastructure of an organisation.

5.2 Future Directions

The Binary String Analysis considered in Section 3 provides the basis for the development of a machine
learning paradigm based on computing a set of statistical metrics associated with the signal given by
Equation (8). This metric set, coupled with other metrics on the spectrum of the signal, for example, is
then used to construct a feature vector which forms the input to an ANN. Whether such an approach will
be out-ranked by a deep learning approach remains to be identified, but either way, it is important to
monitor binary streaming to identify encrypted signatures and structures. In doing so it may be possible
to continually monitor internet traffic to estimate the presumed increase in encrypted communications
as a result of the approach considered in Section 2 growing in popularity.
The methods introduced in this paper are focused on symmetric encryption in which the key and/or EC
cipher are assumed to be exchanged before an encrypted communication is initiated. There are many key
exchange methods that can be used for this purpose which exploit asymmetric encryption or a no keys
exchange protocol using a three-way pass [53]. Both methods incur weaknesses especially with the
evolution of quantum computing that makes legacy systems for key/algorithm exchange based on the
RSA algorithm effectively redundant. Asymmetric cryptographic systems are based on trapdoor functions,
i.e. functions that have a one-way property unless a secret parameter (trapdoor) is known, and to date,

Copyright © Society for Science and Education United Kingdom 53

 Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
no counterpart of a trapdoor transformation is, as yet, known in chaos theory. The use of EC and other
approaches that come under the discipline of AI may be of value in the exploration and possible solutions
to this problem.
The applications considered in Section 4 are illustrative of the way in which AI can be used to generate
encrypted data for authentication using the IoT coupled with a communications unit which is in
widespread use, i.e. the smartphone. The applications presented are typical of the way in which wireless
communications, data security and AI are being coupled to secure the transmission of information.
In addition to the optical cipher technique discussed in Section 4.2, it may be of interest to supplement
reflective (microspheres) powder with Graphene powder, given that mono-layer Graphene absorbs some
2.3% of incident photons at any wavelength. Thus, by applying Graphene powder with a lacquer to a
surface patch of a banknote, for example, it may be possible to use a Graphene based cipher obtained
from the absorbance of light rather than, or in addition to, the reflection of the light by retro-reflective
powder. Alternatively, it may be possible to dope ink using either retro-reflective powder or Graphene
powder subject to the cost effectiveness of such an approach compared to the use of patch ciphers of the
type illustrated in Figure 10.
There are of course numerous other applications of machine learning in cryptography than have been
considered in this paper [78], [79]. Coupled with complementary developments such as advances in DNA
computing [80], deep learning and quantum computing, it can be expected that current encryption
algorithms, protocols and standards associated with information and cyber security will change radically
in the near future. Moreover, these changes will be under the research, development and control of a
much broader global spectrum of society.
Acknowledgements The authors acknowledge the support of the University of KwaZulu-Natal,
Technological University Dublin and the Science Foundation Ireland, University of Western Cape, the
University of Wales and Warsaw University of Technology. The authors are grateful to Adam Magnuski
for his collaboration in regard to the development of the RFID technology discussed in Section 4.1.

REFERENCES

[1] Copeland, B. J., Artificial intelligence. Encyclopedia Britannica, 2020.

https://www.britannica.com/technology/artificial-intelligence

[2] Turing, A. M., Computing Machinery and Intelligence. Mind, 1950, 49: p. 433-460 [Online]
Available from: https://www.csee.umbc.edu/courses/471/papers/turing.pdf [Accessed 27
April 2020].

[3] Wikibooks, Artificial Neural Networks/Neural Network Basics, 2020.

https://en.wikibooks.org/wiki/Artificial_Neural_Networks/Neural_Network_Basics

[4] Kostadinov, S., Understanding the Back-Propagation Algorithm, Towards Data Science, 2019.
https://towardsdatascience.com/understanding-backpropagation-algorithm-7bb3aa2f95fd

URL: http://dx.doi.org/10.14738/tmlai.83.8219 54
Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

[5] Capterra, Artificial Intelligence Software, 2020. https://www.capterra.com/artificial-intelligence-

software/

[6] AI System Inc., 2020. https://aisystems.ai/

[7] MathWorks, AI with MATLAB, 2020. https://uk.mathworks.com/campaigns/offers/ai-with-matlab.html

[8] MathWorks, Machine Learning with MATLAB, 2020.

https://uk.mathworks.com/campaigns/offers/machine-learning-with-matlab.html

[9] Joshi, P., Artificial Intelligence with Python: A Comprehensive Guide to Building Intelligent Apps for
Python Beginners and Developers, Packet Publishing Limited, 2017. ISBN: 978-1-78646-439-
2. https://www.amazon.com/Artificial-Intelligence-Python-ComprehensiveIntelligent/dp/178646439X

[10] Turner, M. J., Blackledge, J. M. and Andrews, P., Fractal Geometry in Digital Imaging, Academic
Press, 1998. ISBN: 0-12-703970-8, 1998.

[11] Fractal Geometry: Mathematical Methods, Algorithms and Applications (Ed. J. M. Blackledge,
A. K.

[12] Evans and M. J. Turner), Woodhead Publishing: Series in Mathematics and Applications, 2002.
ISBN: 190427500.

[13] Deep Learning in Digital Pathology, Global Engage, 2020. http://www.global-engage.com/life-

science/deep-learning-in-digital-pathology/

[14] Google Cloud. AI & Machine Learning Products, Advanced Guide to Inception V3 on Cloud TPU,
https://cloud.google.com/tpu/docs/inception-v3-advanced

[15] Zhang, W., Itoh, K., Tanida, J. and Ichioka, Y., Parallel Distributed Processing Model with Local
Space-invariant Interconnections and its Optical Architecture, Applied Optics, 1990, 29(32), p.
4790–4797. https://drive.google.com/file/d/0B65v6Wo67Tk5ODRzZmhSR29VeDg/view

[16] Blackledge, J. M., Digital Image Processing, Woodhead Publishing Series in Electronic and
Optical Materials, 2005, ISBN-13: 978-1898563495. https://arrow.tudublin.ie/engschelebk/3/

[17] MathWorks, Introducing Deep Learning with MATLAB, 2020.

https://uk.mathworks.com/campaigns/offers/deep-learning-with-matlab.html

[18] Maghrebi, H., Portigliatti, T., Prouf, E., Breaking Cryptographic Implementations Using Deep
Learning Techniques, Security, Privacy, and Applied Cryptography Engineering (SPACE), 6th
International Conference, 2016. [Online] Available from: https://eprint.iacr.org/2016/921.pdf

[19] Bezobrazov, S., Blackledge, J. M. and Tobin, P., Cryptography using Artificial Intelligence, The
International Joint Conference on Neural Networks (IJCNN2015), Killarney, Ireland, 12-17 July,
2015.

Copyright © Society for Science and Education United Kingdom 55

Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
[20] Asiru, O. F., Blackledge, J. M. and Dlamini, M. T., Application of Artificial Intelligence for
Detecting Computing Derived Viruses, 16th European Conference on Cyber Warfare and
Security (ECCWS 2017), 2017, University College Dublin, Dublin June 29-30, p. 647-655.

[21] Hoare, O., Enigma: Code Breaking and the Second World War - The True Story through
Contemporary Documents. 2002, Introduced and Selected by Oliver Hoare, UK Public Records.

[22] Vernam, G. S., Cipher Printing Telegraph Systems for Secret Wire and Radio Telegraphic
Communications, Transactions of the American Institute of Electrical Engineers, 1926, 55, p.
109–115.

[23] Blum, L., Blum, M. and Shub, M. A., Simple Unpredictable Pseudo-Random Number
Generator, SIAM Journal on Computing, 1986, 15 (2), p. 364–383.

[24] Matthews, R., On the Derivation of a ‘Chaotic’ Encryption Algorithm, Cryptologia, 1984, 8(1), p.
29–41.

[25] Ptitsyn, N. V., Deterministic Chaos in Digital Cryptography, PhD Thesis, De Montfort University,
UK and Moscow State Technical University, Russia, 2002.

[26] Erdos, P., Pemereance, C. and Schmutz, E., Carmichael’s Lamba Function, Acta Arithmetica,
1991, 58(4), p. 363-385.

[27] Kocare, L., Chaos and Cryptography, 2001, http://rfic.ucsd.edu/chaos/ws2001/kocarev.pdf.

[28] Blackledge, J. M., Ptitsyn, N. V. and Chernenky, V. M., Deterministic Chaos in Digital
Cryptography, First IMA Conference on Fractal Geometry: Mathematical Methods, Algorithms
and Applications (Eds. J M Blackledge, A K Evans and M Turner), Horwood Publishing Series in
Mathematics and Applications, 2002, p. 189-222.

[29] Blackledge, J. M., Cryptography and Steganography: New Algorithms and Applications, Centre
for Advanced Studies Text-books, Warsaw University of Technology, ISBN: 978-83-61993-05-6,
2012.

[30] Eiben, A. E. and Schoenauer, M., Evolutionary Computing, Information Processing Letters,
2002, 82(1), p. 1-6.

[31] Ehrenberg, R., Software Scientist: With a Little Data, Eureqa Generates Fundamental Laws of
Nature, Science News, 2012, 181(1), p. 20-21.

[32] Rechenberg, I., Evolutionsstrategien, Simulationsmethoden in der Medizin und Biologie.

Springer, 1978, 302, p. 83-114.

[33] Fogel, L. J. Owens, A. J. and Walsh, M. J., Artificial Intelligence through Simulated Evolution,
Wiley, 1966.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 56
Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

[34] Holland, J. H., Adaptation in Natural and Artificial Systems: An Introductory Analysis with
Applications to Biology, Control and Artificial Intelligence. University of Michigan Press, 1975.

[35] Eiben, A. E. and Smith, J. E., Introduction to Evolutionary Computing. Springer, 2003, ISBN 978-
3-662-05094-1.

[36] Nutonian, Eureqa: The AI-Powered Modeling Engine, 2020. [Online] Available from:
www.nutonian.com [Accessed 29 January 2020].

[37] Schmidt, M. and H Lipson, H., Distilling Free-form Natural Laws from Experimental Data,
Science, 2009. [Online] Available from: https://science.sciencemag.org/content/324/5923/81

[38] Dubcakova, R., Eureqa: Software Review. Genetic Programming and Evolvable Machines, 2011,
12(2), p. 173-178.

[39] Blackledge, J. M., Bezobrazov, S. Tobin, P. and Zamora, F., Cryptography using Evolutionary
Computing, Proc. IET ISSC2013, Letterkenny, Co Donegal, Ireland, June 20-21, 2013 (Awarded
Best Paper Prize for ISSC2013). https://arrow.tudublin.ie/aaconmuscon/5/

[40] NIST, Cryptographic Algorithm Validation Program (CAVP), 2020, [Online] Available at
https://csrc.nist.gov/Projects/cryptographic-algorithm-validation-program

[41] NIST, CAVP Random Number Generators, 2020, [Online] Available at

https://csrc.nist.gov/Projects/cryptographic-algorithm-validation-program/Random-Number
Generators

[42] Random.org, 2020. [Online] Available from https://www.random.org/

[43] Araneus Information Systems Oy, 2020. [Online] Available from https://www.araneus.fi/

[44] Ntolo, S., Student No. 3175732, Course Work Assignment Report, Information and Communications
Security (COS 792), Department of Computer Science, University of Western Cape, 2018.

[45] Mosola, N. N., Dlamini, M. T., Blackledge, J. M., Eloff, J. H. P. and Venter, H. S., Chaos-based Encryption
Keys and Neural Key-store for Cloud-hosted Data Confidentiality, Southern Africa Telecommunication
Networks and Applications Conference (SATNAC), 2017, September 3-10, p. 168-173.

[46] Wikipedia, List of Random Number Generators, 2020. [Online] Available from
https://en.wikipedia.org/wiki/List_of_random_number_generators

[47] Blackledge, J. M. and Ptitsyn, N. V., On the Applications of Deterministic Chaos for Encrypting Data on the
Cloud, Third International Conference on Evolving Internet, INTERNET 2011, 19-24 June, IARIA, 2011,
Luxembourg, p. 78-87. ISBN: 978-1-61208-008-6, 2011.

[48] Tobin, P., Blackledge, J. M., Bezobrasov, S., Personalized Encryptor Generation for the Cloud using
Evolutionary Computing, International Workshop on Nonlinear Maps and their Applications (NOMA-
2015), Dublin, Ireland, 15-16 June, 2015.

Copyright © Society for Science and Education United Kingdom 57

Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
[49] Tobin, P., Blackledge, J. M., McKeever, M., Secrecy and Randomness: Encoding Cloud Data Locally using A
One-Time Pad, International Journal on Advances in Security, IARIA, 2017, 10(3), p. 167-181.

[50] Blackledge, J. M., Information Hiding using Stochastic Diffusion for the Covert Transmission of Encrypted
Images, Proc of IET ISSC2010 UCC Cork, 23-24 June, 2010.

[51] Blackledge, J. M. and Al-Rawi, A. R., Steganography using Stochastic Diffusion for the Covert
Communication of Digital Images, IANEG International Journal of Applied Mathematics, 2011, 41(4), p.
270-298.

[52] Blackledge, J. M., Tobin, P. Myeza, J. and Adolfo, C. M., Information Hiding with Data Diffusion Using
Convolutional Encoding for Super-Encryption, Mathematica Aeterna, 2017, 7(4), p. 319 – 356.
https://arrow.tudublin.ie/engscheleart2/127/

[53] Blackledge, J. M., Govere, W. and Sibanda, D., Phase-Only Digital Encryption, IAENG
International Journal of Applied Mathematics, 2019, 49(2), p. 212-228.
https://arrow.tudublin.ie/engscheleart2/192/

[54] Blackledge, J. M., Tobin, P., Govere, W. and Sibanda D., Phase-only Digital Encryption using a
Three-pass Protocol, ISSC2019, IEEE UK and Ireland Signal Processing Chapter, Maynooth
University, 17-18, June 2019.

[55] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A. and
Bengio, Y., Generative Adversarial Networks, Proceedings of the International Conference on
Neural Information Processing Systems (NIPS), 2014, p. 2672–2680.
https://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf

[56] Szilard, L., On the Decrease of Entropy in a Thermodynamics System by the Intervention of
Intelligent Beings, Zeitschrift fur Physik, 1922, 53, p. 840–856.
http://fab.cba.mit.edu/classes/862.19/notes/computation/Szilard-1929.pdf

[57] Shannon, C. E., A Mathematical Theory of Communication, The Bell System Technical Journal,
1949, 27, p. 379–423, 1948.
http://people.math.harvard.edu/~ctm/home/text/others/shannon/entropy/entropy.pdf

[58] A. N. Kolmogorov, A. N., Entropy per Unit Time as a Metric Invariant of Automorphism, Russian
Academy of Sciences, 1959, 124, p. 754-755.

[59] Sinai, Y. G., On the Notion of Entropy of a Dynamical System, Russian Academy of Sciences, 1959,
124, p. 768-771.

[60] Marsaglia, G., Random Numbers Fall Mainly in the Planes, Proc. Natl. Acad. Sci., 1968, 61(1), p.
25–28.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 58
Transactions on Machine Learning and Artificial Intelligence Volume 8, Issue 3, June 2020

[61] Gao, Y., Kontoyiannis, I., Bienenstock, E., Estimating the Entropy of Binary Time Series:
Methodology, Some Theory and a Simulation Study, Entropy, 2008, 10, p. 71-99.

[62] Pincus, S., Approximate Entropy as a Measure of System Complexity, Proc. Natn. Acad. Sci., 1991,
88, p. 2297-2301.

[63] Rukhin, A. L., Approximate Entropy for Testing Randomness, J. Appl. Prob., 2000, 37(1), p. 88-
100.

[64] Richman, J. S. and Moorman, J. R., Physiological Time-series Analysis using Approximate Entropy
and Sample Entropy, American J. of Physiology-Heart and Circulatory Physiology, 2000, 278(6),
p. H2039-H2049.

[65] Chen, W., Zhuang, J., Yu W. and Wang, Z., Measuring Complexity using FuzzyEn, ApEn, and
SampEn, Med. Eng. and Phys., 2009, 31(1), p. 61-68.

[66] Carroll, J. M., The Binary Derivative Test: Noise Filter, Crypto Aid and Random Number Seed
Selector, Simulation, 1989, 53, p. 129-135.

[67] Carroll, J. M. and Sun, Y., The Binary Derivative Test for the Appearance of Randomness and its
use as a Noise Filter, Technical Report No. 221, Department of Computer Science, University of
Western Ontario, 1998.

[68] Bruwer, C. S., Correlation Attacks on Stream Ciphers using Convolutional Codes, Masters Thesis,
University of Pretoria, 2005.

[69] Croll, G. J., BiEntropy – The Approximate Entropy of a Finite Binary String, 2013.
https://pdfs.semanticscholar.org/d8d6/bf07050f9dcd026bfafe62c079b4e309dd7a.pdf

[70] Text to Binary translator, Rapid Tables, 2019. [Online] Available from
https://www.rapidtables.com/convert/number/ascii-to-binary.html

[71] Radio Frequency Identification (RFID), Wikipedia, 2020, https://en.wikipedia.org/wiki/Radio-

frequency_identification

[72] NFC Flex Stamp Antenna, 2020. file://mac/Home/Downloads/L152.pdf

[73] Gołofit, K. and Wieczorek, P. Z., Chaos-based Physical Unclonable Functions, MDPI Applied Sciences,
Special Issue on ‘Side Channel Attacks’, 2019. https://www.mdpi.com/2076-3417/9/5/991

[74] Allen, L. et al., The Hubble Space Telescope Optical Systems Failure Report, NASA, Technical Report,
November, 1990. https://ntrs.nasa.gov/search.jsp?R=19910003124

[75] Wikipedia, Retroreflector, 2020. https://en.wikipedia.org/wiki/Retroreflector

[76] Glowtec, Reflective Powder, 2020. https://glowtec.co.uk/Reflective-powder/

Copyright © Society for Science and Education United Kingdom 59

Jonathan Blackledge and Napo Mosola; Applications of Artificial Intelligence to Cryptography, Transactions on
Machine Learning and Artificial Intelligence, Volume 8 No 3 June, (2020); pp: 21-60
[77] Shor, P.W., Algorithms for Quantum Computation: Discrete Logarithms and Factoring, Proceedings 35th
Annual Symposium on Foundations of Computer Science, 1994, IEEE Comp. Soc. Press, p. 124–134.

[78] Open Quantum Safe Project, Software for Prototyping Quantum Resistant Cryptography, 2016.
https://openquantumsafe.org/

[79] Stamp, M., Introduction to Machine Learning with Applications in Information Security, Chapman &
Hall/CRC Machine Learning & Pattern Recognition, 2018. ISBN-13: 978-1138626782.

[80] Alan, M. M., Applications of Machine Learning in Cryptography: A Survey, 2019.

https://arxiv.org/pdf/1902.04109.pdf

[81] Advances of DNA Computing in Cryptography (Eds. S. Namasudra and G. Chandra Deka), Chapman and
Hall/CRC, 2018.

URL: http://dx.doi.org/10.14738/tmlai.83.8219 60