ALARM: Digital Twin-Assisted Reinforcement LeARning for Anomaly Mitigation in vehicular Edge Computing

As smart vehicles connect to the Internet of Vehicles (IoV), their electronic systems face severe cyberattack risks. Missing a time-sensitive task during an attack can be disastrous. While current defences often rely on static rules to fix system failures, in this paper (ALARM), we introduce an adaptive, Digital Twin (DT)-assisted Reinforcement Learning (RL) framework to detect and mitigate varying threats. To overcome the risks of training and executing AI algorithms on cars in real time, a DT safely pre-trains the RL agent using simulated attacks. During an intrusion, the agent acts as a smart healing module, reallocating tasks to healthy processors in real-time to meet strict deadlines while satisfying resource constraints, ensuring safe vehicle operation. Experiments show that ALARM maintains a Quality of Service (QoS) between 27% and 68% as system utilisation varies from 40% to 90%. The system is highly efficient at detecting threats, reaching 98.49% accuracy even when the signs of an attack are minute. This makes ALARM a powerful tool for making vehicle networks more reliable and energy-efficient, essential for passenger safety in the IoV.