Making a Case for Federated Learning in the Internet of Vehicles and Intelligent Transportation Systems
AACCEPTED FOR PUBLICATION IN IEEE NETWORK MAGAZINE SPECIAL ISSUE MAY 2021 - AI-EMPOWERED MOBILE EDGE COMPUTING IN THE IOV 1
Making a Case for Federated Learning in theInternet of Vehicles and Intelligent TransportationSystems
Dimitrios Michael Manias and Abdallah Shami
Abstract —With the incoming introduction of 5G networksand the advancement in technologies, such as Network Func-tion Virtualization and Software Defined Networking, new andemerging networking technologies and use cases are taking shape.One such technology is the Internet of Vehicles (IoV), whichdescribes an interconnected system of vehicles and infrastructure.Coupled with recent developments in artificial intelligence andmachine learning, the IoV is transformed into an IntelligentTransportation System (ITS). There are, however, several opera-tional considerations that hinder the adoption of ITS systems,including scalability, high availability, and data privacy. Toaddress these challenges, Federated Learning, a collaborativeand distributed intelligence technique, is suggested. Through anITS case study, the ability of a federated model deployed onroadside infrastructure throughout the network to recover fromfaults by leveraging group intelligence while reducing recoverytime and restoring acceptable system performance is highlighted.With a multitude of use cases and benefits, Federated Learningis a key enabler for ITS and is poised to achieve widespreadimplementation in 5G and beyond networks and applications.
I. I
NTRODUCTION
With recent advancements in networking technologies suchas Mobile Edge Computing (MEC) and Network FunctionVirtualization (NFV) and the incoming introduction of FifthGeneration (5G) networks, a multitude of new applicationsand use cases are being realized, such as the Internet ofVehicles (IoV). These up-and-coming use cases require strin-gent Quality of Service (QoS) guarantees and strict ServiceLevel Agreements (SLAs) to ensure network properties suchas scalability, flexibility, elasticity, high availability, and per-formance; however, these properties cannot be realized to theirfull potential without the use of intelligence techniques suchas Machine Learning (ML) and Advanced Analytics (AA). Byleveraging the enabling networking technologies to create theIoV and combining it with intelligence techniques, the IoVtransforms into an Intelligent Transportation System (ITS).
A. Mobile Edge Computing
MEC is a technology that pushes cloud computing servicesto the network edge. By utilizing the network edge, severalbenefits arise, including ultra-low latency, high bandwidth,and real-time applications. Due to these benefits, MEC hasbeen highlighted as an enabling technology for next-generation
Dimitrios Michael Manias and Abdallah Shami are with the Departmentof Electrical and Computer Engineering at the University of Western Ontarioe-mail: { dmanias3, Abdallah.Shami } @uwo.ca networking and applications. Some examples of use casesstemming from MEC implementations include video analytics,location services, IoV and augmented reality [1]. Anothermajor advantage of MEC implementations is the reducedtraffic experienced on core networks, something which isessential given the growing network connectivity demandexperienced globally by Network Service Providers (NSPs).When considering use cases such as the IoV, real-time, ultra-low latency processing available at the edge of the networkis essential for critical services, including collision avoidanceand infotainment services, such as virtual and augmentedreality. However, MEC itself requires the virtualization ofnetwork infrastructure to enable cloud resource utilization atthe network edge. B. Network Function Virtualization
NFV technology was proposed by the European TechnicalStandards Institute in 2012 and entails the abstraction of net-work functions from dedicated hardware [2]. Once abstracted,the network functions are converted to Virtual Network Func-tions (VNFs), which are software-based applications that runon universal hardware such as data center servers and edgeservers. Through NFV technology, several benefits can berealized, such as reduced capital and operational expenditures,improved network performance and operation, and improvednetwork health [3]. When considering NFV technology, oneof the greatest challenges is the management and orchestration(MANO) of VNFs, which includes tasks such as the placementof VNFs on network servers, VNF scaling and VNF migration.The increasing complexity of networks, coupled with the NP-hard computational complexity of these problems [4], hasled NSPs to consider alternate approaches to address NFVMANO.
C. Intelligence and NFV
Recently, the use of intelligence techniques such as MLand AA have been increasingly popular when consideringNFV MANO functionalities. This increase in popularity isattributed in part to a major paradigm shift from analyticalnetwork modelling to data-driven network modelling. Withthe generation of increasing amounts of network data, NSPsare beginning to adopt intelligence technologies that leveragethe previously untapped data and extract meaningful infor-mation. There are several benefits associated with the useof intelligence and the adoption of data-driven modelling in a r X i v : . [ c s . N I] F e b CCEPTED FOR PUBLICATION IN IEEE NETWORK MAGAZINE SPECIAL ISSUE MAY 2021 - AI-EMPOWERED MOBILE EDGE COMPUTING IN THE IOV 2 Fig. 1. Basic Intelligent Transportation System Overview
NFV MANO. Firstly, since network complexity is increasing,analytical system modelling becomes increasingly difficult; bymodelling the system directly from the generated data, NSPscan get accurate system models without the need to describethe system mathematically in its entirety. Additionally, in thecase of NFV MANO functionalities ( i.e.,
VNF Placement),intelligence can be used to learn from past optimal VNFplacements and predict future placements in real-time. Thisability to predict optimal placements enables real-time optimaldecision making, something which was previously not possibledue to the complexity of optimization problem formulationsand the static nature of near-optimal heuristic solutions. Ad-ditionally, the use of intelligence enables a plethora of newand innovative functionalities such as traffic, demand, andlatency prediction and forecasting, which can be used tooptimize scaling operations. Additionally, one of the mainfunctionalities of 5G and beyond networks is automation in theform of self-healing, self-configuration, and self-optimization;all of these functionalities require extensive use of intelligencefor forecasting and prediction thereby making intelligence anessential requirement of all 5G and beyond networks.
D. Intelligent Transportation Systems
The combination of MEC, NFV, and intelligence sets thestage for several emerging use cases, most notably, ITS [5].The ITS framework envisions a system of connected vehiclescommunicating with each other and with Intelligent Infrastruc-ture (II) using dedicated short-range communications. ThroughITS connectivity, Vehicular Clients (VCs) will have accessto multiple types of services, including traffic, emergency,and infotainment [5]. Considering the use of MEC in thissystem, these services will be provided through NSPs bymeans of VNFs placed close to the user at the networkedge. To enable this, several Roadside Units (RSUs), actingas edge servers, will be placed along roads. These unitswill have the capability of collecting information such astraffic and weather conditions, locally processing that data, andsending it to various entities within the ITS system. Figure 1shows a basic ITS system and highlights the various entitiespreviously mentioned (1 – RSU, 2 – VC, 3 – II). Figure 1also highlights the presence of entities capable of performinglocal intelligence, marked with the gears.
E. The Challenges of Machine Learning Implementation inITS Systems
There are several challenges regarding the implementationof ML in highly dynamic environments such as the IoVand ITS. These challenges are categorized into four majorgroupings, system complexity, model performance, privacy, and data management [6]. System complexity is a majorchallenge when considering ITS as it is a very volatileenvironment; while roadside infrastructure may be constantin nature, vehicular clients are continually entering and leav-ing the system. This volatility presents a unique challengeregarding ML implementation as the operational domain iscontinuously changing, something which is not easily handledby traditional ML [6]. The shifting operational domain leadsto the second major challenge being model performance. Asthe functional domain changes, the performance of the modelis severely impacted. Models using static local intelligencecannot adapt to changing environments and can become inef-fective as their performance is severely degraded. Consideringthe criticality of an ITS system, the safeguard of humanlife is paramount. Any level of compromise in the systemranging from its infrastructure to its data, can endanger publicsafety. Finally, with an increasing number of network nodeswith processing capabilities, the management of data becomesincreasingly critical. Since the roadside infrastructure will havelimited resource capacity, special consideration must be maderegarding the efficient storage of data. Since the resourcecapacities of the roadside infrastructure will not allow for thestorage of extensive data sets, the training phase of traditionallocalized ML techniques can be compromised due to the lackof sufficient data.
F. Why Federated Learning?
When considering possible methods of advanced intelli-gence applied to an ITS, Reinforcement Learning (RL) andFederated Learning (FL) are two standout options. RL is anintelligence method capable of learning complex policy deci-sions in a dynamic environment and has the ability to adapt toa changing domain through continual and experiential learning[5]. However, the success of the computationally intensivetraining phase of RL is greatly dependent on its simulatedtraining environment, the design of its reward function, andthe tuning of its hyperparameters [5]. Considering an ITSwith multiple light-weight points of presence having limitedprocessing and storage capabilities, the implementation of RLbecomes increasingly challenging.When considering an ITS, security and privacy areparamount; this applies both to vehicles and the data theygenerate. Since not all intelligence is created equal, a techniquethat maximizes data privacy while still meeting the requiredperformance is essential to the feasibility of such a system.Furthermore, when considering the system environment, adistributed and communication-efficient intelligence methodis required. Additionally, a form of intelligence that canovercome faults and failures quickly, thereby ensuring systemresilience and service continuity, is a major contender forimplementation. To simultaneously address privacy concernswhile ensuring a resilient and intelligent system capable ofexcellent performance, we make a case for FL in the IoVand ITS. The remainder of this paper is organized as follows.Section II Outlines FL, its advantages, and its benefits. SectionIII outlines the applications of FL and showcases possible usecases. Finally, Section IV summarizes and concludes the paper.
CCEPTED FOR PUBLICATION IN IEEE NETWORK MAGAZINE SPECIAL ISSUE MAY 2021 - AI-EMPOWERED MOBILE EDGE COMPUTING IN THE IOV 3
II. F
EDERATED L EARNING
FL is a key enabling technology for many emerging usecases. The following section will outline the main aspectsof this technology, its benefits, and present a case studyillustrating its effectiveness. Table 1 summarizes the variousbenefits provided by Federated Learning implementations, asoutlined throughout this section.
TABLE IF
EDERATED L EARNING B ENEFIT S UMMARY
Benefit DescriptionCollaborativeLearning Enables collaboration between multiple entities forintelligence.DistributedLearning Enables intelligence across distributed nodes andenables intelligence in new and emerging use cases.DecentralizedLearning Local nodes can train their own models and canoperate independently without the aggregation agentin failure scenarios.Data Availability Allows the leveraging insights obtained from massiveamounts of data without the need for locally hostingthe data.Data Privacy Data is collected and processed at local nodes andnot shared globally.Scalability New nodes can easily be integrated into existingsystems.Fault Recovery Offers a rapid recovery scheme for nodes experienc-ing faults.High Availability Offers system continuity during failure scenarios.CommunicationEfficiency Model updates are sent instead of complete modelsor data.
A. What is Federated Learning
FL is a machine learning technique first proposed by Googlein 2017 as a way of providing decentralized and collabora-tive learning across distributed nodes [7]. Initially, FL wasconsidered for applications relating to smartphones such astext prediction; however, since then, it has been used in fieldssuch as medicine and image recognition. The FL architectureconsists of multiple federated nodes and an aggregator agent.Initially, a global model is created by the aggregator agentand is distributed to all the federated nodes. Once received,the nodes begin by training the model on their locally storeddata. Since each node is responsible for the collection and/orstorage of its own data, each federated node possesses a uniquetraining data set. Once the federated nodes have trained theirmodels for several iterations (pre-defined in the aggregationscheme), the nodes compare the initial global model receivedto their locally trained model. An update is generated by eachfederated node, whereby the results of the comparison arestored. It must be noted that this update does not include thelocally trained model itself; instead, it lists the discrepanciesbetween the global and local models, thereby highlightingthe changes made during the training process. The aggregatoragent then samples the nodes according to the aggregationscheme and collected the model updates. Once collected,the updates are used to create a new global model that isdistributed to the nodes, and the process repeats itself. Figure2 outlines the FL process at a high level. However, to fullyappreciate the entirety of the FL process, a more granularanalysis of the federated nodes and the aggregator agent isrequired.
B. Federated Nodes
Federated Nodes are unique and versatile entities capableof data collection and processing, model training, and networkcommunication. When considering federated nodes in IoV andITS scenarios, the main entities which can be classified assuch are VCs, RSUs, and II. In the envisioned ITS scenario,these entities will have various sensors and will be gathering amultitude of data. Each of these entities will have processingand communication capabilities.
C. Aggregator Agent
The aggregator agent is the main orchestrator behind theFL training process and is responsible for determining howoften and how many nodes will be contributing to the globalmodel update based on the aggregation scheme. Developingan aggregation scheme is the main challenge associated withFL as it directly affects the performance of FL process.When considering the aggregation scheme, the first decisionto be taken is how the model updates received from each of thenodes will be consolidated into a single update used to gen-erate the new global model. Traditionally, averaging has beenused to combine all the model updates; however, increasingly,new strategies are emerging which take into considerationindividual properties of the federated nodes, including resourceavailability and node criticality. By creating a more refinedupdate aggregation strategy that captures domain-specific in-formation, model performance can be greatly improved.The second consideration which must be made is thefrequency of model updates. This frequency is a joint consid-eration between the number of local training iterations beforethe update and the time required to locally train the model.This consideration brings to light one of the main attributes ofFL data; traditionally, most ML models assume that a datasetis independent and identically distributed (i.i.d.); however,this assumption is invalid for FL as it operates under theassumption that the data is non-i.i.d. When thinking aboutthis intuitively, especially considering an IoV scenario that thedata collected will not follow the i.i.d. assumption. Considera scenario where a group of federated nodes are responsiblefor training object detection models at given intersections; thebusier intersections will have more local data compared tothose which are located in low-traffic areas, thereby creatinga non-identically distributed scenario. Going back to theaggregation scheme, nodes with more local data will require alonger time to train their models and generate the update. Thisimbalance in data and, therefore, an imbalance in training timemust be taken into consideration when selecting the frequencyof model updates.The final consideration, which must be made regardingthe aggregation scheme, is the sampling strategy. Due to thenature of FL, especially in applications such as the IoV, allnodes will be gathering different data. Performing a modelupdate using an insufficient amount of data is a futile taskas it will not bring an overall benefit to the global model.Conversely, waiting for a node with an excessive amount ofdata to finish its training can hold back the training processand drastically reduce the speed at which the global model
CCEPTED FOR PUBLICATION IN IEEE NETWORK MAGAZINE SPECIAL ISSUE MAY 2021 - AI-EMPOWERED MOBILE EDGE COMPUTING IN THE IOV 4
Entity Process
Aggregation
Agent
Federated
Nodes
Generate Global Model (GM) GM DistributeGM GM Collect Local Data (LD) LD Use LD to TrainLocal Model (LM) LM
Compare GM and LM and Generate Local Update (LU) LU Send LU to Aggregation AgentLU Aggregate all LUs and generate new global model (GM’)
GM'
Has termination criterion been met?
GM = GM’
FL Process Terminatedyesno GM FL Process Initiated
Fig. 2. Federated Learning High-Level Process Map
Aggregation Scheme
AveragingN1
N2 N3
𝑁1 + 𝑁2 + 𝑁33
AveragingN1 N2 N3 𝑤1 ∗ 𝑁1 + 𝑤2 ∗ 𝑁2 + 𝑤3 ∗ 𝑁3𝑤1 + 𝑤2 + 𝑤3 w1 w2 w3 N1 Single Local Training EpochMultiple Local Training Epochs N1 Update
Update
All Nodes Selected for UpdateSome Nodes Selected for UpdateBasic AveragingDomain-Based Weighted Averaging Fig. 3. Aggregation Agent Scheme Considerations is trained. Furthermore, due to the dynamic nature of thenetwork supporting MEC implementations, there could bea high volume of traffic resulting in higher communicationlatency at a given node; using resources to communicate modelupdates at such a time can further burden the network andhave adverse effects on QoS requirements and service delivery.Finally, in the case of a fault or failure, a particular node cango offline; waiting for a model update at this time would beexclusively dependent on the extent of the fault and the abilityto recover from the fault, and can greatly reduce the globalmodel’s ability to progress in training. Taking these scenariosinto consideration, a sampling of nodes must be conductedto determine which nodes will provide the greatest benefit tothe current global model through their updates. This decisionis domain-specific as in certain situations, the nodes with thegreatest amount of data provide the most benefit, whereas, inother situations, the nodes with a moderate amount of datacan speed up the training process and converge to a solutionfaster. Figure 3 outlines the various aspects of the aggregationscheme previously mentioned.
D. What are the Advantages of Federated Learning?
The main advantage of FL, arising from its decentralizedand collaborative learning properties, is the preservation ofprivacy during the model training process. Since only localmodel updates are sent to the aggregator agent, the data usedto train the local models remains with the local node. Thefact that neither the aggregator agent nor the other federatednodes have access to a given node’s individual data unlocksan incredible potential for FL to be used in privacy-sensitiveapplications. One such privacy-sensitive application which hasalready begun to implement FL is the intelligent healthcaresector [8] . The most prevalent example in this sector iscollaborative model training between hospitals. Traditionally,due to data privacy and patient confidentiality concerns, hos-pitals had only had access to their local patient data whentraining models, which created a data-availability bottleneckand was a key deterrent for the use of ML since the localdata was insufficient. With FL, several hospitals can contributeto the training of a global model, which leverages updatesfrom all their local models, thereby enabling an individualhospital to use insights from other participating hospitals toimprove the performance of their models. This collaborationleads to the second main advantage of FL, data availability.
CCEPTED FOR PUBLICATION IN IEEE NETWORK MAGAZINE SPECIAL ISSUE MAY 2021 - AI-EMPOWERED MOBILE EDGE COMPUTING IN THE IOV 5
Through the use of FL, not only is it possible to ensuredata privacy, but it is also possible to use insights from non-local data, thereby increasing the total amount of data usedfor model training. This is especially important for nodeswith low amounts of local data as an inadequate amount ofdata can prove to be detrimental to a model’s performance.Additionally, the method by which this data is processed isextremely advantageous as all nodes are solely responsible forprocessing their local data. The ability to leverage global databut only process local data makes FL especially applicable toresource-constrained nodes with low processing capabilities,such as those found in MEC IoV systems.Another advantage of FL is the communication efficienciesexperienced during the model training. Since there is no datatransfer between local nodes and the aggregation agent, and nocommunication between nodes, FL is very efficient comparedto centralized collaborative learning strategies. When consider-ing the complex IoV and ITS network layout, this communi-cation efficiency can alleviate a potential strain on networkresources during times of high traffic, something which isincreasingly important considering the critical services usingthe network and its services.The final key advantage of FL is seen in a fault or failuremitigation scenario. In the case that a local node goes offlineor loses connectivity, the FL training process can still continuenormally. In the case where a node loses connectivity, it stillhas its local model to use and, therefore, won’t experience afault. However, in the case where a node goes offline andloses its local model, once communication is restored, theaggregator agent can quickly push the current global modelto the node. Having a global model the node can resumeoperation and restore performance. Furthermore, having aglobal model and then applying local training is an incrediblyefficient failure mitigation strategy. Firstly, the model hasalready been developed and trained through a number ofiterations. Secondly, local data from the failed node con-tributed to model development before the failure. Consideringthe abovementioned points, as soon as the node receives theglobal mode, it can begin using it, and through a minimalnumber of local training iterations, pre-failure performancecan be quickly restored. The effectiveness of this mitigationstrategy is highlighted when compared to alternative schemes,such as complete model retraining and the reinstatement ofhistorical models. The complete model retraining strategy,as the name suggests, requires the entire retraining of themode from scratchl using local data, something which isvery time efficient and can increase downtime. The historicalmodel reinstatement firstly requires the storage of previousmodels, and secondly, the retraining of these models, whichis an improvement to the complete model retraining strategy,however, can prove to be inefficient in a highly dynamic andvolatile network scenario. Since the global model is ‘live’ andconstantly updating based on incoming data, it is the preferredfailure mitigation strategy. To illustrate the capabilities of FLand its advantage over other failure strategies, we present acase study.
E. Case StudyA Federated Learning – enabled RSU is located in a lowtraffic area and performs object classification tasks. Due to arecent construction project, the neighbourhood is experiencinga variety of previously unseen traffic. The RSU continuallytrains its local model to adapt to the changing environment;however, it experiences an outage and loses its local model.It can select one of the previously outline failure recoverystrategies. Which one should it choose?
To simulate the scenario described above, the MNIST digitdataset [9] was used. Initially, a model is trained on a setnumber of digits ( i.e., even), which simulates the operation ofthe RSU before the construction project. Once the constructionproject begins, and previously unseen traffic is experienced,odd digits are gradually introduced into the training and testingdatasets; however, training of this model is not complete as thefault is experienced. There are three possible fault recoveryscenarios outlined:1. A model is retrained from scratch using the local dataavailable. The training and testing sets of this model willconsist of even and odd integers to simulate the moment atwhich the fault was experienced.2. A previous model is re-instated, and then training occursusing local data. The dataset used to train this model is thesame as scenario 1.3. A global federated model trained using 10 system nodesrepresenting RSUs spread throughout the network is pushed tothe failed node, and is instantly used to resume performance.A comparison between the three mentioned failure mitiga-tion strategies is exhibited in Figure 4. As seen through thisfigure, the federated recovery strategy performs the best asit instantly restores normal operation and performance returnsto acceptable levels without any additional training iterations. The main advantages of FL described throughout this sectionencompass many operational benefits, which are summarizedin Table 1. III. A
PPLICATIONS
The following section outlines some of the various applica-tions of FL in the IoV and ITS, including RSU Intelligence,NFV Orchestration, and Vehicular Intelligence.
A. RSU Intelligence
As previously mentioned, ITS have entities known as RSUs,which will be located along roadways. These RSUs willbe equipped with sensors to collect data and will possessbasic processing capabilities. When considering the generalFL application architectures, RSUs perfectly match since theyare repeated entities capable of data collection and processing.Since the RSUs will be collecting (and receiving) a varietyof different data, they are capable of applying FL in manydifferent scenarios. One of the most important and prevalentapplications is image processing. When considering a systemof fully autonomous vehicles, image processing is essential code for this use case is available at https://github.com/Western-OC2-Lab/FL-IOV-ITS.git CCEPTED FOR PUBLICATION IN IEEE NETWORK MAGAZINE SPECIAL ISSUE MAY 2021 - AI-EMPOWERED MOBILE EDGE COMPUTING IN THE IOV 6
Fig. 4. Failure Recovery Strategy Comparisson both onboard the vehicular clients as well as roadside throughthe RSUs. Image processing tasks on these entities can rangefrom pedestrian detection to collision reporting. When consid-ering any transportation system, there are differing levels ofvehicular traffic experienced throughout the system; this trafficimbalance can put certain RSUs at a great disadvantage asthe data collection in low traffic areas will minimal comparedto the data collection in high traffic areas. To mitigate this,FL applied across all RSUs can enable the collective use ofall RSU data and the distribution of a complete model toRSUs which otherwise wouldn’t have an adequate amount ofdata to train a local model of their own. Another advantageof using FL for image processing is the ability to leveragethe differing conditions ( i.e., lighting) in RSU-collected im-ages. Some RSUs will be placed in fully lit areas, whileothers might be placed in neighbourhoods with many treesand shaded regions. To ensure proper object detection, allRSUs should have models capable of operating under varyingand changing conditions. Considering the criticality of objectdetection applications in RSUs, FL is essential for ensuringmodel performance and, subsequently, the safety of driversand pedestrians alike.
B. NFV Orchestration
NFV Orchestration presents a very appealing and intriguingapplication for FL. The most enticing property of FL, makingit a candidate for NFV Orchestration is its privacy and securitydue to the critical services ( i.e., financial, emergency) NFV-enabled networks support. When considering the IoV andITS as an extension of current networks, NFV Orchestra-tion will play a critical role in vehicular service delivery.However, there is one major difference between traditionalNFV Orchestration and NFV Orchestration for MEC enabledIoV and ITS. With many delay-sensitive applications rangingfrom vehicular safety and routing to immersive virtual realityservices, NSPs are facing an unprecedented challenge ofreducing end-to-end application latencies to less than 1msfor ultra-low latency applications [10]. To do this, points ofpresence such as RSUs which are placed at the very edge of the network will act as network nodes capable of hosting VNFs.However, without proper Orchestration, the stringent latencyrequirements will not be met. Traditionally, NSPs have had toresort to using near-optimal heuristic solutions to address theinfeasibility of optimization problem formulations due to theirruntime complexity, however, with the adoption of ultra-lowlatency applications, near-optimal heuristic solutions are alsobecoming infeasible. To mitigate the infeasibility of traditionalNFV Orchestration techniques, NSPs have been exploring MLas an alternative. Already, ML techniques are being appliedto NFV Orchestration tasks such as VNF placement [11],[4] and migration [12] with incredible success due to theirability to approach the performance of the optimization modelformulations with an incredible reduction in time-complexity.When considering ML in the IoV and ITS, the use of FL isthe next logical step. Due to the increasing types of services,each with its own complexities coupled with the expansionof traditional networks to MEC enabled networks, NFV Or-chestration activities will drastically change. The transition toMEC-enabled networks will usher in the vast expansion ofnetwork nodes. This will require the partitioning of currentnetworks into much smaller sub-networks, each with their ownNFV Orchestrator. Additionally, incorporating aspects from5G networking, each of these sub-networks may be furtherpartitioned using network slicing techniques. This increasedcomplexity promotes the use of FL as NFV Orchestrators fromdifferent network partitions can use collaborative ML trainingto create models capable of completing orchestration taskssuch as VNF Placement, Scaling, Termination and Migration.Figure 5 illustrates the complexity of vehicular requests in a5G-enabled ITS.
C. Vehicular Intelligence
Vehicular intelligence in the IoV and ITS can describe aplethora of applications, including in-vehicle intelligence ( i.e., communications), image processing ( i.e., lane detection), andforecasting ( i.e., road conditions); however, a very interestinguse case currently being heavily explored in the manufacturing
CCEPTED FOR PUBLICATION IN IEEE NETWORK MAGAZINE SPECIAL ISSUE MAY 2021 - AI-EMPOWERED MOBILE EDGE COMPUTING IN THE IOV 7
Requested Service Provider 1:Vehicular Application Requested Service Provider 2:Infotainment Application
Requested Service Provider 1:
Infotainment Application Communication Application Requested Service Provider 2:Medical ApplicationInfotainment Application Requested Service Provider 1:Infotainment Application
Requested Service Provider 2:
Vehicular Application
Requested Service Provider 1:
Infotainment ApplicationRequested Service Provider 2:Communication Application
Service Provider 1 Network Service Provider 2 NetworkMobile Broadband SliceMassive IoT SliceMission Critical SliceInfotainment Slice Mobile Broadband SliceMassive IoT SliceMission Critical SliceInfotainment Slice
Fig. 5. Vehicular Application Requests in 5G-enabled ITS sector relates to predictive maintenance. Predictive mainte-nance uses operational data to predict when a specific compo-nent will fail and suggests a maintenance schedule that willpro-actively address the failure through planned maintenance.Unplanned maintenance is extremely costly in manufacturingas a break in production can result in the loss of millions inrevenue. Additionally, reactive maintenance takes a large tollon assets and reduces their lifespan. Several implementationsof predictive maintenance models have shown incredible suc-cess; one main example stemming from the oil and gas indus-try is Royal Shell Corp who has prevented millions of dollarsin lost revenue and damages while improving the longevityof their assets [13]. The same benefits can be experiencedwhen considering the application of predictive maintenanceon vehicular clients in the IoV and ITS. It is estimated that in2016 there were 1.32B vehicles globally [14], a number whichis expected to grow exponentially in the next decades. In anITS scenario, each of these vehicles would have an extensivelog capturing real-time measurements from its various sensors.By leveraging FL, along with this vast collection of data,comprehensive predictive maintenance models can be built.While the mechanics of how to select which data is used tobuild these models is still an open question ( i.e., brand-based,location-based, network-based), the impact FL will have onpredictive vehicular maintenance is indisputable.IV. S
UMMARY AND C ONCLUSION
As demonstrated throughout this paper, Federated Learninghas incredible potential in terms of its applicability to theInternet of Vehicles and Intelligent Transportation Systems.From the various number of use cases, including RoadsideUnit Intelligence, Network Function Virtualization Manage-ment and Orchestration, and Vehicular Intelligence to theincredible number of benefits it provides, Federated Learningis a key enabler for next-generation networking technologies.R
IEEE Network , vol. 28, no. 6, pp. 18-26, 2014.[4] D. M. Manias et al. , “Machine Learning for Performance-Aware VirtualNetwork Function Placement,”
IEEE GlobeCom , Waikoloa, USA, 2019,pp. 1-6.[5]
Intelligent Transport Systems (ITS); Vehicular Communications; BasicSet of Applications; Definitions et al. , ”Federated learning: Strategies for improving commu-nication efficiency.” arXiv preprint, 2016.[8] T. S. Brisimi, et al. , ”Federated learning of predictive models fromfederated electronic health records.”
International journal of medicalinformatics pp. 59-67, 2018.[9] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. ”Gradient-based learningapplied to document recognition.”
Proceedings of the IEEE in IEEE Journal on Selected Areas in Communications , vol. 37,no. 3, pp. 643-655, March 2019.[12] D. M. Manias, H. Hawilo, and A. Shami,“A Machine Learning - BasedMigration Strategy for Virtual Network Function Instances,”