Study On Vehicular Networks Security Technology Based On Reinforcement Learning

There is traffic-sensitive information such as traffic control information,accident warnings,and vehicle positioning.Malicious users launch attacks to block,eavesdrop or temper this information,which may lead to privacy leakage,communication interruption or even loss of control of the vehicle,causing serious economic losses and social impact.Among them,the jammer may block legal communication in the vehicular network,and even cause a denial of service attack.Message falsification attacks destroy the integrity and authenticity of the data.Eavesdroppers aim to steal private information from vehicular network users.Therefore,the thesis studies the vehicular network communication security scheme based on reinforcement learning to suppress the attack motive,achieve active defense,and resist multiple attacks such as jamming attacks,message falsification attacks,and eavesdropping attacks.In this paper,we propose a reputation-based security mechanism based on indirect reciprocity to regulate the behaviors of on-road units and suppress the motivation of attacks.For the unpredictable attack strategy scenario,we also propose a communication mode selection strategy based on deep reinforcement learning for an on-road unit to optimize the performance of packet delivery rate and attack rate.The paper further proposes an improved algorithm based on deep Q-network,which improves the communication optimization rate of the on-road unit that can support the deep learning algorithm,and further suppresses the attack motive of the selfish users in the large-scale vehicular network.For example,the proposed Deep Q-network based schemes can reduce the attack rate by 48.9%compared with the reinforcement learning based strategy with 8 OBUs.A cooperative power control scheme for roadside unit broadcast is proposed to resist jamming attack.By applying reinforcement learning techniques,we propose a deep reinforcement learning based anti-jamming scheme,in which an optimal communication policy can be achieved without the need to know the jamming model.To accelerate the learning process,this scheme compresses the state space observed by convolutional neural networks,and improves the update speed of state-action value and reduce storage space by memory module.Simulation results show that the cooperative power control scheme based on reinforcement learning increases the packed delivery rate by 12%compared with an independent power control scheme with 3 roadside units.