Machine Translated by Google

AI@EDGE: A Secure and Reusable Artificial

Intelligence Platform for Edge Computing
Roberto Riggioÿ , Estefanía Coronado† , Neiva Linder‡ , Adzic Jovanka§ , Gianpiero Mastinú ,
Leonardo Gorattiÿÿ, Miguel Rosa††, Hans Schotten‡‡, and Marco Pistorex

*RISE Research Institutes of Sweden AB, Stockholm, Sweden; [email protected]

† i2CAT Foundation, Barcelona, Spain; Email: [email protected]
‡Ericsson AB, Stockholm, Sweden; Email: [email protected]
§Telecom Italia SpA, Turin, Italy; Email: [email protected]
Politecnico di Milano, Milano, Italy; Email: [email protected]
ÿÿSafran Passenger Innovations, Wessling, Germany; Email: [email protected]
††Aerotools, Madrid, Spain; Email: [email protected]
‡‡German Research Center for Artificial Intelligence, Kaiserslautern, Germany; Email: [email protected]
x
Fondazione Bruno Kessler, Trento, Italy; Email: [email protected]

Abstract—Artificial Intelligence (AI) has become a major become a major innovative force in almost any vertical, and
innovative force and a major pillar in the fourth industrial revolution.
are being foreseen as one of the pillars boosting the fourth
This trend has been recognized by the European Commission, industrial revolution.
who has pointed out how high-performance, intelligent, and
secure networks are fundamental to the evolution of the multi- This projection is supported by the fact that cloud and
service Next Generation Internet (NGI). While great progress mobile networks are already converging at several techno
has been done in the accuracy and performance of AI-enabled logical levels. From one side, cloud technologies, such as
platforms, their integration in autonomous decision-making
cloud-native and virtualization, are making their way into the
and critical systems requires end-to-end quality assurance.
AI@EDGE addresses these challenges harnessing the concept telecom operators' domain. At the same time, networking use
of “reusable, secure, and trustworthy AI for network cases and requirements, such as access-aware operations
automation”. To this end, AI@EDGE targets significant and service function chaining, are influencing the evolution of
breakthroughs in two fields: (i) general-purpose frameworks cloud technologies. Despite such advances, the cloud
for closed-loop network automation capable of supporting
operators' approach remains centralized with few large data
flexible and programmable pipelines for the creation, utilization,
and adaptation of the secure, reusable, and trustworthy AI/ML centers deployed at key locations, while telecom operators
models; and (ii) converged connect-compute platform for manage a distributed infrastructure based on a hybrid multi
creating and managing resilient, elastic, and secure end-to-end cloud approach.
slices supporting a diverse range of AI-enabled network
applications. Cooperative perception for vehicular networks, In this context, 5G is a paradigm shift: its high performance
secure, multi-stakeholder AI for Industrial Internet of Things, in terms of latency, bitrate, and reliability, call for a
aerial infrastructure inspections, and in flight entertainment are technological and business convergence between the cloud
the uses cases targeted by AI@EDGE to maximize its commercial, computing
societal, and
andenvironmental impact. 5G features like slicing,
the telecom worlds.
Index Terms—AI, 5G, MEC, automation, disaggregated RANs, Multi-access Edge Computing (MEC), and more flexible radio
ML-based security, hardware acceleration, serverless platforms
connectivity can be used to support qualitatively different
applications, and to deliver a richer user experience, faster
I. INTRODUCTION interactions, large scale data processing, and machine to
Artificial Intelligence (AI) systems are irreversibly set on the machine communications . Nevertheless, the challenges to
evolutionary path of every future vertical as well as of every be overcome to realize this connectivity/computing
object and service we human will interact with in the near convergence are still notable. In particular, the increasing
future. This trend is motivated by the need to support elastic number of control and optimization dimensions of the end-to-
and demanding real-world use cases such as automated end 5G infrastructure may result in an overly complex network
mobility, e-health, gaming, etc. In this scenario, it is recognized that operators and vendors may find it difficult to operate, manage, and evo
that the operators will have the opportunity to fill a central role AI and Machine Learning (ML) technologies will be crucial
in providing innovative solutions for application and service in the cloud-network convergence process and will help
developers that want to combine the advanced capabilities of operators achieve a higher level of automation, increase
5G with the fluid cloud-based application development network performance, and decrease the time-to-market of new features.
processes emerged in the last decade, such as, for example, Early attempts at applying AI/ML in the cellular domain can
the Platform as the Service (PaaS) and the microservice/ be found in several academic works [2], [3], [4], [5].
serverless models. A significant example of this ecosystem However, it cannot be expected that each and every
are AI-enabled applications, which have subsystem of future access, edge, core, and cloud segments will
Machine Translated by Google

Fig. 1: Functional overview of the AI@EDGE platform.

employ distinct and separate AI tools and datasets. Such an A. Network automation platform leveraging flexible and
approach would lead to AI-silos, slowing down advances vital reusable AI pipelines
to achieving sustainable networking and ultra-scale complex
services relying on distributed compute-connect fabric.
The approach of AI@EDGE to answer the above mentioned 5G is a full paradigm shift where high performance in terms
challenges has two lines of action. First, we will design, of latency, bitrate, and reliability, calls for a technological and
prototype, and validate a network and service automation business convergence between cloud computing and network
platform able to support flexible and programmable pipelines working. The increasing number of control and optimization
for the creation, utilization, and adaptation of secure and dimensions of the 5G infrastructure may lead to an overly
privacy-aware AI/ML models. Second, we will use this platform complex network that operators and vendors may find it
to orchestrate AI-enabled end-to-end applications. difficult to operate, manage, and evolve. AI technologies will
Here, we introduce the novel concept of Artificial Intelligence be key in this roadmap and, although early attempts to address
Functions (AIFs) to refer to the AI-enabled end-to-end this issue can be already found [6], [3], [7], [8], they focus on
applications sub-components that can be deployed across the specific problems that cannot be extrapolated to other network
AI@EDGE platform. Finally, the AI@EDGE platform will be segments . However, in a full automated system, having
validated using four well-chosen use cases with specific distinct and independent AI tools and datasets would make
requirements that cannot be satisfied by current 5G networks impossible the sustainable management of networking and
according to 3GPP R15 and 3GPP R16, in particular in terms highly-scalable services on a distributed connect-compute
of support for latency-sensitive and highly dynamic AI enabled platform.
applications.
Therefore, the challenge is to implement a general-purpose
The rest of the paper is structured as follows. In Sec. II we
network automation framework capable of supporting flexible
discuss the challenges addressed by the AI@EDGE Project.
and reusable end-to-end AI pipelines. Scalable AI/ML models,
Section III covers the AI@EDGE concept for beyond 5G
fast data pipelines and effective data dissemination models
networks. The four reference use cases are described in Sec.
are crucial to realize automation at scale. Some of the main
IV. Finally, Sec. V concludes this paper.
pillars of AI@EDGE are the potential of multi-access edge
II. CHALLENGES com putting and the powerful mechanisms of scalable
The main objective of AI@EDGE is to build a secure distributed and federated learning in the 5G context. Based
connect-compute platform capable of enabling the automated on this, AI@EDGE addresses this challenge by developing a
roll-out and management of large scale heterogeneous edge platform for closed-loop automation that allows the deployment
and cloud computing infrastructures. To this end, the platform of AI/ML compute infrastructures over the edge, which also
encompasses the required APIs to enable the deployment of accounts for secure isolation of co-located AI/ML algorithms
large-scale virtual compute overlays (eg, containers and by multiple stakeholders. running on shared MEC resources.
serverless instances, etc.) across a multi-connected The progress on this challenge will enable two main results:
heterogeneous infrastructure able to support a range of future
critical applications. This is illustrated in Fig. 1, which • Scaling of AI/ML distributed algorithms to ensure
represents the functional overview of the platform across two application performance and model reliability under
dimensions: (i) AI/ML-driven multi-connected applications at varying resource availability.
massive scale; and (ii) AI/ML-centered security. However, • Zero-touch end-to-end network and service management
this am bitious objective involves several important challenges. including the creation, utilization, and adaptation of
Such challenges are described in detail in the following subsections. reusable AI/ML pipelines in a connect-compute platform.
Machine Translated by Google

Migration and Scaling Migration and Scaling Compute, storage, and

slice caching resources

Latency-critical AIFs Low-latency AIFs Latency-tolerant AIFs

hardware acceleration

AI-driven platform management and control function

non-RT RAN Intelligent
Fault Management Resource Management security management
Controller (RIC)

Multi-connected computing infrastructure interface layer

DU DU
UC
UK near edge Far Edge

DU
DU DU UC
UC BBUs
UK near edge Far Edge central ocean

UC
UK DU

near edge

UK

End-user Processing Far/Near Edge Central Cloud

level processing level processing level

Fig. 2: The AI@EDGE AI-enabled connect-compute platform.

B. Secure and resilient ML for multi-stakeholder environment This challenging research question will mainly lead to three key results:
lies

From a security perspective, there are several relevant aspects for the • Increase in attack detection speed and intrusion resistance, including
success of 5G and beyond systems. In production networks, the potential early detection and automated configuration and speedup of
risks for tangible assets, such as servers and human beings victims of countermeasures. • Model propagation and
attacks, is significantly increasing. Consequently, intrusion detection computational efficiency to optimize the set of exchanged parameters,
systems are a native part of the current 5G security architecture. eg, the weights of a neural network, and their aggregation methods.
• Privacy security of the model parameters exchanged between
However, they are usually under proprietary licenses, which highlights the edge devices.
need to enable an open exchange of models and parameters for intrusion
detection, especially in multi stakeholder environments. It must be noted
that this issue worsens when the platforms are driven by AI/ML models,
III. THE AI@EDGE CONNECT-COMPUTE FABRIC FOR BEYOND
since a new and dangerous attack surface is added. The ability to achieve
5G NETWORKS
resilience and service continuity requires simple and information-effective
data-driven models, suitably designed for running on Internet of Things
and MEC devices with limited resources. Therefore, the challenge is to The design of the AI@EDGE platform envisions the auto mated roll-
provide light-weight, secure and resilient ML systems that are robust to out of adaptive and secure compute overlays, and a new generation of AI-
evasion and poisoning attacks. enabled end-to-end applications. Such applications are made possible by
AI@EDGE through the introduction of the novel concept of AIFs, which
refer to the AI-enabled applications sub-components that can be deployed
With the advent of ML, privacy techniques have been recently revisited and chained across the various levels of the architecture. This vision is
presented in Fig. 2, in which AI@EDGE combines a set of cutting-edge
to better accommodate the trade-off be tween privacy risk and data
usefulness in the construction of ML pipelines. Federated Learning (FL) cloud computing and 5G concepts with a reusable, secure, and privacy
[9], [10], [11] and adversarial networks [12] have recently been used to preserving AI/ML layer to enable an innovative network automation
cover the security aspects. FL allows assembling a common model platform supporting all aspects of network and service management
combining local models built from edge devices without disclosing any including the deployment and scaling of AIFs of different nature (ie, latency-
data. However, this approach poses numerous problems such as local critical, low-latency, and latency-tolerant AIFs) over a distributed facility,
biases, temporal offsets, etc. Security is one of the cornerstones of the and the various tasks needed to deploy such applications, eg the creation
architecture of AI@EDGE. of a new network slice.

Consequently, AI@EDGE aims to define and implement an ML

methodology following a distributed paradigm, allowing the development, The remainder of this section describes the technological enablers on
implementation and evaluation of effective intrusion detection algorithmic which AI@EDGE builds to convert the connect compute platform into a
framework, accounting for security and isolation of co-located AI/ML reality, and that compose the functional blocks of the AI-driven platform,
algorithms. tackling as sketched in Fig. 3.
Machine Translated by Google

Fig. 3: Conceptual architecture and functional blocks of AI@EDGE.

A. Distributed and decentralized connect-compute platform domains. Defining these AI-enabled applications involves the
Enabler. AI@EDGE combines the Function as a Service representation of their AIFs (ie, AI-enabled applications
subcomponents). To this end, AI@EDGE leverages standard
(FaaS) paradigm with serverless computing, hardware
knowledge representation languages and well-known state of-
acceleration (GPU, FPGA and CPU), and a cross layer, multi
the-art ontology engineering methodologies, to represent AIFs,
connectivity-enabled disaggregated Radio Access Network (RAN)
as well as their relationships and status at the different ent levels
into a single connect-compute platform to allow over the-top
of the technology stack. Conversely, AI@EDGE considers “de-
providers to fully use the 5G capabilities though well established
facto” standards for cloud and edge services orchestration, and
cloud-native paradigms to develop and run applications. The
from their emerging variants (eg, FaaS) to for the end-to-end
serverless and FaaS approaches are gaining attention as cloud
orchestration and chaining of AIFs.
computing models in which the infrastructure provider manages
Innovation. AI@EDGE aims to build on the above on
on-demand infrastructure and resources, while the stakeholders
technologies to propose a reference model that provides the
(eg, service providers) can focus only on their core activities.
tools to describe AIFs, their requirements (eg, storage, hardware
Building on this, AI@EDGE intends to define a set of open APIs,
acceleration, etc.), and the necessary metadata for their or
by which network operators, vertical industries, service providers
chestration, which will also compose a catalog of available AIFs.
and users can interact with the network on a neutral host model.
Furthermore, AI@EDGE envisions innovative solutions for end-
Furthermore, the platform encompasses the path towards a
to-end orchestration to partition AIFs across different segments
hybrid multi-cloud native deployment supporting Virtual Machines
attending to their requirements (as shown in Fig. 2) considering
(VMs) and containers and their integration with the serverless
the heterogeneity and complexity of the underlying edge
paradigm.
Innovation. AI@EDGE aims to account for this mixture by computing platforms, and the collection of valuable quality of
service indicators to create complex AI-enabled applications and
extending the current ETSI MEC/NFV architectures with
detect abnormal situations.
application and application-intent models able to capture the
huge heterogeneity in the application building domain. For this C. Hardware-accelerated serverless platform for AI/ML
purpose, AI@EDGE takes as a basis the Cloud Native Application Enabler. The most recent hardware acceleration solutions
Bundling (CNAB) initiative, and will propose its extension to (FPGA, GPU, and CPU) and privacy preserving ML tech niques
support for serverless technologies (besides VMs and containers). allow and speedup the execution of sensitive and computing-
In addition, context and metadata from application and application- intensive workloads over the same platform. The deployment of
intent modeling studies are meant to be used for realizing heterogeneous acceleration platforms at the edge enables
intelligent control and management of applications and services advanced processing scenarios to be exploited in far more
deployed over the serverless decentralized and distributed complex processing functions. In addition to GPUs, which are
AI@EDGE platform. This can be observed for a wide range of currently the prime solution for AI/ML processes acceleration,
verticals in Fig. 2. FPGAs are gaining momentum for deployments at the edge due
to their ability to ensure optimal performance to execute
B. Orchestration of Artificial Intelligence Functions
specialized functions (eg, real-time network intensive processing),
Enabler. The provisioning of AI-enabled applications over a in an energy and cost efficient manner.
distributed computing platform requires reference models and Innovation. AI@EDGE makes resource-aware hardware
standards, especially in heterogeneous and complex scenarios acceleration techniques a key point of its design with the goal of
as edge computing platforms spanning across multiple increasing resource efficiency across the computing
Machine Translated by Google

continuum. As depicted in Fig. 2, AI@EDGE aims to go a traffic simulator like Car Maker (or VI-Grade) and to design,
step further by exploring approaches to tame accelerators' implement, and test the digital twinning of a mix of real and
heterogeneity and enable their integration (both GPUs and emulated vehicles. The goal is to recreate the network-level
FPGAs) with the serverless computing concept, in order to data exchange required to build a cooperative perception be
offer a unified platform able to allocate resources and migrate tween emulated vehicles and a virtual human-driven vehicle.
functionality between accelerators on different edge devices The AI-based digital twinning process will make use of the
or between edge and cloud infrastructures. AI-enabled application features and of the distributed and
centralized serverless AI@EDGE platform. A key role here
D. Cross-layer, multi-connected, disaggregated radio access will be the AI@EDGE network and service automation
Enabler. Supporting beyond 5G use cases requires lying features allowing the digital twinning to cope with mutating
on different communication technologies to increase reliability, radio network environment. The AI@EDGE platform will be
as exposed in Rel15 and Rel16 through dual connectivity interfaced with a 5G network emulator to allow testing a
techniques using data duplication at the PDCP layer. However, broader range of scenarios and network configuration and
besides reliability, various use cases demand greater flexibility related 5G Key Performance Indicators (KPIs) will be
and openness in the RAN to implement more advance multi- measured.
connectivity layers and and to enable a higher degree of
automation. This demand is being promoted in O RAN B. UC2: Secure and resilient orchestration of large Industrial
specifications [13], which propose an open architecture where Internet of Things (IIoT) networks
RAN control and management functions are divided into near- Smart factories will be characterized by 5G connectivity
Real-Time (nRT) and non-Real-Time (nonRT) RAN Intelligent using massive machine-type communications slices to
Controller (RIC) . AI@EDGE aims to rely on both multi- interconnect both IIoT and IoT devices. Such deployments
connectivity options and O-RAN specifications to deliver a consider the interconnection of independent network
flexible, open and unified platform including 3GPP and segments, potentially managed by different stakeholders.
non-3GPP radio access technologies. Therefore, guaran teeing the confidentiality of proprietary
Innovation. Building on current multi-connectivity options information in a multi stakeholder environment while exploiting
for increased reliability (ie, PDCP data duplication), AI@EDGE as much information as possible in AI/ML detection and
seeks to investigate different approaches for user plane data decision-making is one of the mandatory requirements and
replication over a multi-path with non-3GPP interfaces in key challenges. Furthermore, IIoT environments are very
order to meet the requirements of highly demanding services, sensitive to latency and transmission timing and are governed
as illustrated in Fig. 2. Furthermore, the current O-RAN by strict access control policies, a constraint with which AI/ML solutions mu
architecture will be extended to account for both 3GPP and Conversely, local anomaly detection solutions introduce a
non-3GPP technologies. This will make it possible the significant reduction of the detection capabilities, as local/
collection of dual RAN telemetry, and the extension of the edge detection mechanisms are unable to detect system-wide
network automation platform, which will allow the interaction events and correlations. Federated learning, detection model
with the non-RT and nRT RICs and performing actions on propagation and parameter exchange among edge devices
path selection and switching, among others. can be applied to mitigate these issues and exploit the full
potential of distributed architectures in terms of security and
IV. USE CASES intrusion detection while enforcing data confidentiality between
AI@EDGE will be validated using four high-impact use the stakeholders.
cases. This section provides a brief description of each of This use case envisions the design and validation of
them with a particular focus on commercial relevance and KPIs. mecha nisms for secure orchestration of large scale IIoT
applications on the AI@EDGE platform with the aim of
A. UC1: Virtual validation of vehicle cooperative perception
conducting autonomous workload management on a unified
The automotive use case is cooperative perception. Several connect compute fabric. Flexible, intelligent, and secure
vehicles exchange data related to their trajectories. The data management solutions will be developed with a focus on AI-
are used to build a high-definition map of the surrounding enabled multi-tier infrastructures. Data-driven management
environment that can be used to predict potential collisions. components will be designed to operate in synergy with
Today validation of vehicles' cooperative perception is a security technologies, thereby implementing intelligent,
challenge because cooperative perception deals with protected, and trustworthy network services supporting advanced 5G appli
numerous vehicles that have to: detect in real time the
surrounding traffic scenario; exchange their sensed data; and C. UC3: Edge AI assisted monitoring of linear infrastructures
share their intended maneuvers with other vehicles. Large using drones in BVLOS operation
tests are needed even to address one single traffic scenario. Drones are gaining attention in the industrial world for
Cooperative perception tests become even more complex capturing data in a flexible and innovative way, offering new
when dealing with mixed traffic scenarios. To overcome the operation and maintenance procedures. Complex scenarios
problem of simulating the human behavior by means of a such as monitoring large linear infrastructures can benefit
mathematical model, we plan to interconnect a dynamic from extensive and diverse data capture mechanisms,
driving simulator operated by a real human driver with a configuration in Beyond Visual Line of Sight (BVLOS) flight, low-cost
Machine Translated by Google

operation, and use of combined on-board sensor solutions. V. CONCLUSIONS

Existing studies estimate that a small drone fleet could easily In this paper we present the challenges and conceptual
create 150 terabytes of data per day. To manage this volume architecture of the AI@EDGE project. The project aims to
of data in a production setting drones require a combination of converge and evolve AI/ML and 5G at the network edges with
on board and on the ground edge-computing capabilities. By the goal of providing a flexible platform on top of which the next
leveraging 5G fast data transmission and advanced AI-enabled generation AI-enabled application and services can be
edge image and video processing solution, AI@EDGE will deployed. The paper also describes the four reference use
boost the efficiency of drones in their tasks. cases that will be used to validate the project concept, namely:
To optimize these operations and provide innovative func cooperative perception for vehicular networks, secure, multi-
tionalities a reliable communication technology supporting high- stakeholder AI for Industrial Internet of Things, aerial
rate data transfer and multi-connectivity is required in BVLOS infrastructure inspections, and in-flight entertainment.
scenarios. The 5G network must therefore allow for dynamic AI-
ACKNOWLEDGMENTS
aided workload deployment in combined drone, edge, and
cloud environments, considering the tight latency restrictions This work has been performed in the framework of the
and diverse computational capacities, to process a constant European Union's H2020 project AI@EDGE co-funded by the
dataflow. The coordinated workflow will allow reducing response EU under grant agreement No 101015922. The views expressed
time and quick decision making. Furthermore, AI models will are those of the authors and do not necessarily represent the
also carry on automated incident detection. project. The Commission is not liable for any use that may be
In-flight problem determination as well as the use of drone made of any of the information contained therein.
fleets will significantly benefit from this approach to speed up The authors would also like to acknowledge CERCA
the monitoring process and ultimately support more rapid and Programme/Generalitat de Catalunya for sponsoring part of
frequent preventive actions for maintenance. this work.

REFERENCES

[1] FD Calabrese, P. Frank, E. Ghadimi, U. Challita, and P. Soldati, “En

D. UC4: Smart content & data curation for in-flight entertainment hancing RAN performance with AI,” ERICSSON, Technology Review,
services Jan 2020.
[2] KB Letaief, W. Chen, Y. Shi, J. Zhang, and YA Zhang, “The roadmap to
6g: Ai empowered wireless networks,” IEEE Communications Magazine,
In Flight Entertainment and Connectivity (IFEC) focuses on vol. 57, no. 8, p. 84–90, 2019.
delivering entertainment and engagement to airline passengers. [3] K. David and H. Berndt, “6G Vision and Requirements: Is There Any Need
for Beyond 5G?” IEEE Vehicular Technology Magazine, Vol. 13, no. 3,
The traditional IFEC system made of a central content server p. 72–80, 2018.
and embedded seat screens was conceived to offer only static [4] E. Calvanese Strinati, S. Barbarossa, JL Gonzalez-Jimenez, D. Ktenas,
content over a resource-constrained daisy-chained wired N. Cassiau, L. Maret, and C. Dehos, “6g: The next frontier: From
holographic messaging to artificial intelligence using subterahertz and
network. Currently, it is relentlessly evolving towards a visible light communication,” IEEE Vehicular Technology Magazine, vol.
broadband connectivity platform through on-board Wi-Fi 14, no. 3, p. 42–50, 2019.
technology and ground networks via high throughput satellite [5] F. Tariq, MRA Khandaker, KK Wong, MA Imran, M. Bennis, and M.
Debbah, “A speculative study on 6g,” IEEE Wireless Communications,
connectivity. Recently, with the maturity of 5G, network vol. 27, no. 4, p. 118–125, 2020.
softwarization and AI, the IFEC system has the potential to [6] KB Letaief, W. Chen, Y. Shi, J. Zhang, and YA Zhang, “The Roadmap to
transform into a smart edge-cloud platform composed of 6G: AI Empowered Wireless Networks,” IEEE Communications
Magazine, vol. 57, no. 8, p. 84–90, 2019.
servers, multi-radio access technologies (5G, Wi-Fi, etc.) and [7] E. Calvanese Strinati, S. Barbarossa, JL Gonzalez-Jimenez, D. Ktenas,
embedded seat screens. As the content consumption pattern N. Cassiau, L. Maret, and C. Dehos, “6G: The Next Frontier: From
in the IFEC market is changing rapidly, there is an expectation Holographic Messaging to Artificial Intelligence Using Subterahertz and
Visible Light Communication,” IEEE Vehicular Technology Magazine,
to provide dynamically curated content to the airline passengers, vol. 14, no. 3, p. 42–50, 2019.
based on data about the flight routes and passenger demo [8] F. Tariq, MRA Khandaker, KK Wong, MA Imran, M. Bennis, and M.
graphics (eg, frequent flyer subscriptions). These content Debbah, “A Speculative Study on 6G,” IEEE Wireless Communications,
vol. 27, no. 4, p. 118–125, 2020.
sources include traditional options (major movie makers) and [9] A. Abeshu and N. Chilamkurti, “Deep Learning: The Frontier for Distributed
new sources such as live and near-live streams including news, Attack Detection in Fog-to-Things Computing,” IEEE Communications
viral clips, etc. Magazine, vol. 56, no. 2, p. 169–175, 2018.
[10] WYB Lim, NC Luong, DT Hoang, Y. Jiao, YC Liang, Q. Yang, D. Niyato,
The clear challenge is thus to develop efficient algorithms and C. Miao, “Federated Learning in Mobile Edge Networks: A
and flexible protocols for managing both the content and the Comprehensive Survey,” IEEE Communications Surveys Tutorials, vol.
22, no. 3, p. 2031–2063, 2020.
edge infrastructure. In this context, the lightweight AI@EDGE [11] TD Nguyen, S. Marchal, M. Miettinen, H. Fereidooni, N. Asokan, and A.
platform will equip the IFEC system with edge computing Sadeghi, “D¨IoT: A Federated Self-learning Anomaly Detection System
capabilities and dynamic workload distribution within the for IoT,” in 2019 IEEE 39th International Conference on Distributed
Computing Systems (ICDCS), 2019, p. 756–767.
infrastructure. Furthermore, the multi-connectivity environment [12] A. Tripathy, Y. Wang, and P. Ishwar, “Privacy-preserving adversarial
will allow performing as many tasks as possible at the aircraft networks,” in Proc. of IEEE Allerton, Monticello, IL, USA, 2019.
edge without costly satellite technologies which, together with [13] O-RAN Alliance, “O-RAN Architecture Description v1.0,” February
2020.
AI-aided efficient content distribution, stands for a clear path
forward for the IFEC industry.