Fake face detection through

GAN-generated images of human

faces
Team
B5:
K. Bhanu Sai [20071A6606]
Raghavendra [20071A6620]
K. Santhosh Kumar [20071A6633]
M. Sanath [20071A6659] Guide:
V. Vinod Naik J. Pushpa Kumari
[Assistant Professor, CSE-
AIML]
Agenda

1. Abstract
2. Introduction
3. Existing System
4. Proposed System
5. System Architecture
6. Literature Survey
7. Requirements
8. Implementation
9. Results
10. UML diagrams
11. Conclusion
Abstract
● In the era of advanced deep learning models, particularly with the proliferation of Generative Adversarial
Networks (GANs) generating highly realistic images led to advancements that have spurred exciting
developments in various domains, but they have also given rise to new challenges, particularly in the
realm of digital content authenticity and security.

● This project aims to develop a robust and efficient system for detecting fake human faces generated by
GANs. The proposed system leverages the power of deep learning and explores various techniques to
differentiate between real and GAN-generated faces.

● It involves collecting a diverse dataset of real and GAN-generated face images for training and
evaluation. We employ convolutional neural networks (CNNs) as the core of our detection model,
enabling the system to learn intricate patterns and features associated with real human faces.
Introducti
on
● One of the most promising approaches for generating images of human faces is the use of generative
adversarial networks(GANs). GANs are a type of deep learning model that can be used to generate new
images that are indistinguishable from real images.

● Making use of the wealth of visual information available in face images, these models can learn subtle
patterns, textures, and characteristics unique to real faces.

● Fake face detection has become a pressing issue, as malicious actors exploit these AI-generated images to
create deceptive and misleading content. From the propagation of fake news to identity theft and privacy
breaches, the consequences of such synthetic media can be far-reaching and damaging.

● In this context, deep learning models have emerged as powerful tools to tackle the intricacies of
distinguishing real human faces from their GAN-generated counterparts. By analyzing a range of
parameters, such as facial landmarks, lighting conditions, symmetry, and texture, they can make accurate
distinctions between real and fake faces.
Existing System

Image Forensics: Analyzing image artifacts and inconsistencies introduced during the GAN generation
process.

Feature-Based Methods: Extracting specific facial features and comparing them to real faces to
identify discrepancies.

Deep Learning Models: Training classifiers to distinguish between real and GAN-generated faces
based on learned patterns and features.

Different GAN Networks: GLF-Net for global-local facial network, Gram-net for global image texture
representations, Hierarchical attention memory network (HAMN) to learn and store long-term
dependencies in the input images.
Proposed System
The project aims to develop a robust and efficient system that can ultimately classify an image of
human face as fake or real.
•The proposed system integrates the soft attention mechanism into different CNN architectures and
improves the efficiency of the model to predict whether an image containing a human face is real or
fake.
•Previously, some deep learning models were developed for detecting fake faces. The novelty of our
work is adding an attention layer to the CNN architectures and training them.
• The soft attention layer produces three matrices: value, key, and query, which help assign weights to
all the feature maps extracted by previous layers of CNN architecture.
•The regions that have lower weights are eliminated, and higher-weighted regions are considered for
the classification of images as real or fake. This mechanism helps to focus on relevant areas that
contribute to deciding the exact target label of the photo.
System Architecture
 Literature Survey
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the with DOI addressed research mentioned paper (methods
journal discussed)

1. Global–Local sensors https://doi. Old methods are Proposing a global– 1)Local branch The paper proposes a
org/10.339 model-specific, and the local facial fusion detects the forged novel deepfake
Facial Fusion 0/s2302061 performance is network, namely traces from the detection framework
Based GAN 6 deteriorated when faced GLFNet, to fully facial parts, such called GLFNet.
with more complicated exploit the local as the iris and The framework is able
Generated Face images.They physiological and pupils. to effectively detect
Fake Face proposed a new global receptive GAN-generated fake
network for fake face features. 2)branch adopts a faces that are robust to
Detection detection residual different types of
connection to perturbations.
distinguish real
images from fake
ones 3)Metrics
used are
precision,recall,a
ccuracy
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research mentioned paper (methods
discussed)

2. Global IEEE/CVF https://doi. a.CNN-based a.Analyze the Proposed a new Proposes a novel
Texture Conference org/10.485 fake face relationship between architecture method for detecting
Enhancement on 50/arXiv.2 detectors are not texture and fake face coined as deepfake images
for Fake Face Computer 002.00133 robust to image detection Gram-Net, using global texture
Detection In Vision and editing and b.Propose a novel which leverages information. The
the Wild Pattern cannot generalize global texture global image authors argue that
Recognition well to enhancement texture texture information is
(CVPR) cross-GAN method that is representations more robust to image
images specifically for robust fake distortions and more
b.Texture designed for fake image detection. invariant for face
information is face detection images from different
more robust to c.Show that the GANs than other
image distortions proposed method features, such as local
and more can be effectively edges and shapes.
invariant for face applied to a variety However, CNN-based
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research mentioned paper (methods
discussed)

3 Fake Image IEEE https://doi. Existing deepfake The objective of the The focus of the The paper proposes a
ieeecompu detection methods are research mentioned in paper is to propose a
Detection based on spatial the title is to investigate novel deepfake
method to detect fake
tersociety.
org/10.110 domain features, the potential of the deep detection method images that are
Using An which are not fake detection method that uses an
9/MCSoC generated by artificial
Ensemble of effective on fake that is based on ensemble of CNN
57363.202 intelligence (AI) using an
CNN Models images that are individual face part models specialized
2.00021 manipulated only images, such as eye, for individual face
ensemble of convolutional
Specialized neural
some part of facial cornea, nose, mouth, and parts, such as eye,
For images. face. cornea, nose, mouth, network (CNN) models
The paper propose a and face.
Individual that focus on different
novel deep fake The paper claim
Face Parts detection method that that this method can face parts, such as eye,
uses an ensemble of improve the
CNN models detection accuracy cornea, nose, mouth,
specialized for of deepfakes and face.
individual face parts, compared to existing
which can improve methods that use the
the detection entire facial image
accuracy and reduce
the computational
cost.
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research paper (methods
mentioned discussed)

4. One-Shot 2019 IEEE https://doi. a.As the quality a.Decrease the a.Scene a.Proposed a novel
GAN International org/10.485 of fake faces dependency on understanding One-Shot GAN generated
Generated Symposium 50/arXiv.2 increases, the training data to and object fake face method based
Fake Face on Circuits 003.12244 trained models detect the fake detection to test on out-of-context object
Detection and Systems become Faces. the existence of detection.
(ISCAS) inefficient to out-of-context
detect the novel b.Robust objects b.Instead of using face
fake faces, one-shot learning in the fake analysis approaches, used
training data is method to faces. scene
considered overcome the understanding methods
outdated. problems of b.Bag of words and observed that, some
outdated data and sparse out-of-context
b.Traditional AI and scarcity of vector objects are appeared in
works fine in data. representation fake faces which helped
case of huge to
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research mentioned paper (methods
discussed)

5. Detection of IEEE https://doi. a.Existing a.Develop a deepf a.Hierarchical Proposes a novel

Fake and Transactions org/10.110 deepfake detection ake detection method attention memory deepfake detection
Fraudulent on information 9/TIFS.20 methods are often that is more robust to network (HAMN) method that uses a
Faces via forensics and 20.304776 focused on unseen manipulations, to hierarchical attention
Neural security 8 detecting specific image editing, and learn and store memory network
Memory types of adversarial examples long-term (HAMN) to learn and
Networks deepfakes, such as than existing methods. dependencies in store long-term
face swaps or lip the input images. dependencies in the
syncs. b.Develop a deepfake input images. This
detection method that b.Two-stream makes the network
b.Existing is more efficient to architecture to more robust to unseen
deepfake detection train and deploy than capture both local manipulations, image
methods often existing methods. and global editing, and adversarial
require a lot of features of the examples. The paper
training data, input image. also proposes atwo-
which can be stream architecture to
expensive and capture both local and
time-consuming to global features of the
collect. input image.
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research mentioned paper (methods
discussed)

6. GANprintR: IEEE Journal https://doi. a.GAN-generated a.Develop a new a.A new loss Proposes a new GAN-
of selected org/10.110 deepfakes are GAN-based deepfake function that is based deepfake
Improved Fakes topics in signal becoming
9/JSTSP.2 generation method designed to generation method
and Evaluation processing. 020.30072 increasingly called GANprintR that generate deepfakes called GANprintR and a
of the State of realistic and
50 generates more that are more new large-scale and
difficult to detect.
the Art in Face realistic and difficult to detect. high-quality dataset for
Manipulation b.There is a lack of
sophisticated evaluating deepfake
Detection large-scale and deepfakes than b.A new GAN detection methods
high-quality existing methods. architecture that is called GANprintR-
datasets for more stable and FDDB.It generates
evaluating deepfake b.Develop a new efficient than more realistic and
detection methods. large-scale and high- existing sophisticated deepfakes
quality dataset for architectures. than existing methods
evaluating deepfake c.A new by incorporating a
detection methods post-processing number of new
called GANprintR- step that removes techniques, including a
FDDB. artifacts from new GAN architecture,
generated a new loss function, and
deepfakes. a new post-processing
step.
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research mentioned paper (methods
discussed)

7. On the IEEE/CVF https://doi. 1)Existing 1)Develop a deepfake 1)A deep learning Proposes a novel deep
Detection of Conference org/10.485 deepfake detection detection method that model: The deep learning-based method
Digital Face on Computer 50/arXiv.1 methods often learning model is for detecting deepfake
is more robust to
Manipulation Vision and 910.01717 require a lot of trained on a large images. The method is
Pattern training data, which unseen dataset of real and more robust to unseen
Recognition can be expensive manipulations, image deepfake images. manipulations, image
(CVPR) and time-consuming editing, and 2)A data editing, and adversarial
to collect. This adversarial examples augmentation examples than existing
makes them than existing technique: The methods.
difficult to deploy methods. data augmentation Additionally, the
in technique is used method can be trained
resource-constrai 2)Develop a new data to generate more on a smaller dataset
ned environments. augmentation training data for than existing methods
the deep learning without sacrificing
technique that can be
model. performance.
used to generate
more training data
for deepfake
detection models.
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research mentioned paper (methods
discussed)

8. Classification Internationa https://doi. 1)Existing 1)Develop a deep 1)Feature The DNN is designed
of Real and l Journal of org/10.485 deepfake detection learning-based extraction network: to learn the subtle
Academic 50/arXiv.2 methods are often method for detecting This network differences between real
Fake Human Engineering 202.07145 limited to detecting deepfake images that extracts features and deepfake images,
Faces Using Research specific types of is more robust to from the input even when the
Deep Learning (IJAER) deepfakes, such as unseen images. deepfakes have been
face swaps or lip manipulations, image The features are edited or are adversarial
syncs. This makes editing, and extracted using a examples.To detect
them vulnerable to adversarial examples convolutional deepfake images in new
new types of than existing neural network images, the new image
deepfakes that are methods. (CNN). is passed to the DNN,
constantly being 2)Evaluate the 2)Classification and the DNN outputs a
developed. proposed method on network: This score that indicates the
several public network classifies probability that the
datasets and show the extracted image is a deepfake.
that it outperforms features as real or
state-of-the-art deepfake.
methods in terms of
detection accuracy,
generalization, and
robustness.
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research mentioned paper (methods
discussed)

9. GAN-generat Computer https://doi. 1)GAN-generated 1)Providing a 1)FaceForensics: Provides a

ed Faces Vision and org/10.485 faces are becoming comprehensive This method uses a comprehensive
Pattern 50/arXiv.2 increasingly survey of existing deep neural overview of the state of
Detection: A Recognition 202.07145 realistic and GAN-face detection network to learn the art in
Survey and difficult to detect. methods. the features of GAN-face detection. It
New 2)There is a lack of 2)Identifying the GAN-generated describes a variety of
Perspectives understanding of specific challenges of faces. The network methods for detecting
the specific detecting is trained on a GAN-generated faces,
challenges of GAN-generated dataset of real and including deep learning-
detecting faces. GAN-generated based, physical-based,
GAN-generated 3)Proposing new faces. and physiological-based
faces and directions for 2)DeepFakeHunt methods. The paper
large-scale and research in er and Mesonet also discusses the
high-quality GAN-face detection. challenges of detecting
datasets for GAN-generated faces
evaluating GAN- and proposes new
face directions for future
detection methods research.
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research mentioned paper (methods
discussed)

10. Exposing IEEE Open https://doi. a.Existing a.Develop a deepfake a.Content feature The content and trace
Fake Faces Access org/10.110 deepfake detection detection method that is extraction: The features are then
Through Deep Journal 9/ACCES methods are often more robust to image authors use a combined and fed into a
Neural S.2021.31 focused on editing and adversarial pre-trained deep neural network
Networks 10859 detecting specific examples than existing ResNet-50 model to
with a large receptive
methods extract content
Combining types of field. The deep neural
features from the
Content and deepfakes, such as b.The authors want to image or video. network classifies the
Trace Feature face swaps or lip propose a novel image or video as real
Extractors syncs deepfake detection b.Trace feature or fake.The proposed
method that combines extraction: method was evaluated
b.Previous content and trace feature The trace feature on several public
methods are not extractors. extractor is designed datasets and showed
robust to image to detect traces of that it outperforms
editing and can be the original image or state-of-the-art methods
easily fooled by video that may be in terms of detection
left behind by the
adversarial accuracy,
deepfake generation
examples generalization, and
process.
robustness.
S.NO Title of Title of Reference Problem or Gap Objective of the Focus of the Summary
the paper the journal with DOI addressed research paper (methods
mentioned discussed)

11. GAN-Generate IEEE https://doi The paper did not 1)To develop a 1)The The paper then proposes
d Image JOURNAL .org/10.11 consider the impact GAN-generated self-attention a GAN-generated
Detection With OF of post-processing image mechanism can image detection
Self-Attention SELECTED detection algorithm that uses a
09/JSTSP. on the performance be used to exploit
Mechanism TOPICS IN algorithm that self-attention
Against GAN SIGNAL 2020.2994 of the proposed the structural mechanism. The
uses a
Generator PROCESSIN 523 algorithm. defect in GAN algorithm first uses the
self-attention
Defect G Post-processing is mechanism.
generators. self-attention
a common 2)To evaluate mechanism to learn the
technique used to the proposed 2)The proposed relationships between
improve the quality algorithm on a algorithm different parts of an
number of outperforms other image. Then, it uses this
of GAN-generated
datasets and state-of-the-art information to classify
images. It is not the image as real or
show that it algorithms for
clear how the fake.
outperforms GAN-generated
proposed algorithm
other image detection.
would perform on state-of-the-art
images that have algorithms. 3)The proposed
been
algorithm is
post-processed. 3)To explore
the use of robust to different
self-attention types of GAN
mechanism for generators.
S.NO Title of Title of Reference Problem or Gap Objective of the Focus of the Summary
the paper the journal with DOI addressed research paper (methods
mentioned discussed)

12. A Robust IEEE https://doi 1)The lack of 1)To address the 1)The The paper proposes a
GAN-Generate TRANSACT .org/10.11 robustness to challenges of CBAM-enhance new method for the
d Face IONS ON 09/TCSV post-processing post-processing d Xception detection of
CIRCUITS
Detection operations. and different model is a GAN-generated face
AND T.2021.31
Method Based SYSTEMS 16679 color spaces in variant of the images. The method is
2)The lack of the detection of Xception model based on the use of two
on Dual-Color FOR VIDEO
robustness to GAN-generated that has been feature extraction
Spaces and an TECHNOLO
GY different color face images. enhanced with a models, one for each of
Improved
spaces. convolutional the RGB and YCbCr
Xception 2)To evaluate the
3)The lack of a block attention color spaces. The
proposed module features extracted from
method on a (CBAM). the two models are then
comprehensive comprehensive
evaluation. fused to form a final
set of datasets 2)The final feature vector, which is
and demonstrate feature vector is then classified as either a
its effectiveness. then classified real or a fake face
using a support image.
3)To discuss the
vector machine
potential
(SVM) classifier.
applications of
the proposed
S.NO Title of the Title of Reference Problem or Gap Objective of the Focus of the Summary
paper the journal with DOI addressed research mentioned paper (methods
discussed)

13 Detection of 2018 IEEE https://doi. Existing fakeIn this paper, they The proposed The deep learning
study method is based on methods outperform the
GAN-generate d Conference on org/10.110 image detection the performance a convolutional
Multimedia 9/MIPR.20 methods are not conventional methods in
Fake Images over of several image
Information 18.00084 effective on neural network detecting fake images
forgery detectors
Social Networks Processing and compressed images, against (CNN) that is generated by GANs,
which is common image-to-image trained on a dataset especially when the
Retrieval
for images uploaded translation, both in of both real and
(MIPR) images are
to social ideal conditions, and fake images, both
in the presence of of which have been compressed.The
networks. compression, conventional methods are
compressed. The
routinely performed CNN is trained to more sensitive to
The paper
upon uploading
on social learn the features compression than the deep
proposes a novel learning methods, and
networks. that are
characteristic of their accuracy drops
deep
fake images, even significantly when the
learning-based images are compressed.
when they have
method for been compressed.
detecting fake
images that is
robust to
Requirements

Hardware Requirements:
●Processor: Intel Core i5 or higher (or equivalent AMD processor)
●RAM: At least 8 GB or higher
●GPU: NVIDIA GPU with CUDA support (optional but recommended for faster training)
●Hard disk space: At least 50 GB for storing training data and trained models
●Additionally, we may want to consider using a cloud-based platform such as Google Colab or

Kaggle
Software Requirements:
●Python 3.6 or higher
●TensorFlow 2.0 or higher (for building and training the CNN model)
●Scikit-image (for image processing)
●Matplotlib (for data visualization)
●For deploying web frameworks like Flask to build the backend.
Implementation
● Collected flickr dataset from kaggle that contains real and fake human face images separately.
● Data Augumentation has done to improve the quality of dataset for better results.
● The resulting dataset is splitted into two classes, one is for training and other is to validate the
model.
● We trained different deep learning models using pre-trained models like Vgg16, Inception
ResNet v2, ResNet50, Densnet201 with our dataset that contains over 6800 images.
● Google Colab platform is used for training the model.
● We added soft attention layer into the CNN architecture and train a model that helps to focus
on potential sub-regions that are capable of finding whether the human face is fake or real.
● Then tested the trained model with test data.
Results
An overview of the metrics for all the models with and without a soft attention mechanism can be found
in the tables below. Inception ResNet v2 models gave higher accuracy and precision compared to other
models . Compared to models trained without a soft attention mechanism, those trained using it
produced better results.
Home
Page
Result For Fake Face
Result For Real Face
Use case diagram
Class diagram
Activity Diagram
Conclusion
Thank
you