Detecting and Mitigating Colluding Attacks in Connected Vehicles using Reinforcement Learning

As the vehicles are interconnected and share infrastructure information among each other, the adoption of connected vehicles (CVs) continues to surge. However, CV’s introduce vulnerabilities related to the integrity of the shared data, which may be compromised by either malicious attacks or sensor failures. Detecting these vulnerabilities in CVs has become paramount to provide safety to the passengers and pedestrians in the network. Trust and reputation-based reinforcement learning (RL) algorithms are one way for the detection and handle these vulnerabilities. These algorithms fail to work when the number of vehicles collude together to disrupt the CV network. In response to this, we propose a trust and reputation based RL along with multi level dempster shafer technique to deal with colluding attacks in CVs. This integration involves fusing reputation-based trust management on a vehicle level with a RL agent running on road side unit. We conduct performance analysis on different RL algorithms namely deep Q-networks, actor-critic and proximal policy optimization.