Blockchain-based Robust Federated Learning Scheme In C-V2X Internet Of Vehicles

With the growing scale of connected vehicles in 5th-Generation(5G)mobile communication,traditional cloud-centric approaches can no longer meet the demands of low-latency services.Mobile Edge Computing(MEC)brings the computing platform down from the centralized cloud to the edge of the network closer to the vehicle,and combined with Cellular Vehicle-to-Everything(C-V2X)communication technology can further meet the low latency demand of Internet of Vehicles.On the other hand,a series of intelligent services can be realized by using Artificial Intelligence(AI)technology based on the fast-growing huge amount of data in Internet of Vehicles.Federated learning,as a new distributed machine learning method,can convert the massive data shared between vehicle devices and edge servers into model parameters,effectively solving the problems of data privacy,transmission delay,and resource storage.However,the application of traditional Vanilla Federated Learning(VFL)in the Internet of Vehicles scenario faces some challenges.For example,VFL is vulnerable to local model poisoning attacks,which affects the accuracy of the global model.Also,VFL lacks a reward mechanism,and vehicle nodes with large data samples may be reluctant to share learning results with other nodes.Recent studies have shown that blockchain,as a new distributed data storage system,can be used in conjunction with federated learning to facilitate secure and mutually trusted data sharing in MEC Internet of Vehicles scenarios.Therefore,this thesis proposes a blockchain-based robust federated learning scheme in C-V2 X Internet of Vehicles(BC＿FL),which firstly investigates the aspects of BC＿FL to cope with local model poisoning attacks and reward mechanism design,and finally applies the scheme to help vehicles in resource optimization.The research work in the thesis is divided into three main areas.1.To address the problem that global models are vulnerable to local model poisoning attacks when federated learning is applied in the connected vehicle scenario,a local model accuracy verification scheme based on Practical Byzantine Fault Tolerance(PBFT)is designed and proposed for selecting and storing legitimate local models to guarantee global model performance.After the vehicle uploads the local model to the roadside unit,the roadside unit runs the scheme as the validation node,and the steps are as follows:(1)realize the validation and voting of the local model according to the local model validation voting mechanism;(2)consider the existence of Byzantine validation nodes,and reach the consistency of voting results based on PBFT consensus;(3)judge the legitimacy of the local model based on the consistency of voting results,and store the legal local model is stored into the blockchain.Finally,each roadside unit downloads the legal local model from the blockchain for global aggregation and generates a global model for distribution to vehicles.The simulation results show that the verification scheme can help BC＿FL select the legitimate local model for aggregation,and thus effectively counteract the local model poisoning attack.When 15% of faulty vehicle nodes are present,the global model accuracy of BC＿FL reaches88%,which is 7.5 times higher than that of VFL.2.To address the problem of lack of reward mechanism for federated learning,we design and propose a BC＿FL block generation scheme based on the Proof of Stake(PoS)mechanism,which consists of two parts: reward mechanism and block generation mechanism.First,the reward mechanism is designed according to the different work tasks of BC＿FL nodes,and the rewards are allocated according to them,and the rewards are recorded in the block transaction,and the faulty vehicle nodes and Byzantine validation nodes that cannot perform their work tasks normally will be restricted from participating in BC＿FL,which makes it difficult to obtain continuous rewards;then,by checking the reward records in the block transaction,the roadside unit with the more accumulated stake is randomly selected as the miner node,and the legitimate local model and related voting information and reward records are packaged to generate the block.Simulation results show that this scheme can effectively distribute rewards for vehicles and roadside units in BC＿FL,and the block generation latency is significantly lower than that of the Proof of Work(Po W)mechanism.3.The BC＿FL scheme is applied to the distributed Multi-Agent Deep Q Network(MADQN)based on Internet of Vehicles resource optimization.First,vehicle nodes as intelligence train DQN models based on local Channel State Information(CSI)and upload them to the roadside unit;then the legitimate local models from legitimate vehicle nodes are selected by BC＿FL for global aggregation to generate global models and distributed to intelligence for resource management to maximize the system channel capacity.Simulation results show that the system channel capacity is effectively increased after excluding the faulty vehicle nodes using the BC＿FL scheme.