FlagThis

Abstract

The Internet of Vehicles (IoV) introduces a dynamic and adversarial security environment, where attackers adapt their strategies in response to defensive actions. Most existing intrusion detection systems for IoV rely on offline datasets and static classifiers, which fail to capture sequential decision-making, attacker adaptation, and uncertainty inherent in real-world deployments. In this work, we formulate IoV security as a sequential attacker-defender interaction and model defense as a reinforcement learning problem under partial observability. We propose Quantum Belief-Integrated Reinforcement Defense (Q-BIRD), a quantum-inspired belief representation encoding defender uncertainty about attackers hidden intent using amplitude-based states, enabling non-Bayesian belief evolution under ambiguous observations. The belief state is integrated into a Proximal Policy Optimization (PPO) defender that selects cost-aware mitigation actions such as alerting, throttling, and isolation. We evaluate the approach in a simulated IoV security environment with an adaptive attacker that strategically probes, attacks, and evades detection. The experimental results demonstrate improvements in commonly and widely reported metrics, achieving the reduction in cumulative mean damage, damage variance, attack success rate (ASR), and increased survival probability by 60.4%, 90.2%, 50.0%, 46.4%, respectively. Compared to PPO with classical Bayesian belief, damage variance reduction and ASR improved by 10.2 times and 50%, respectively. These improvements are attributed to Quantum-inspired belief representation, as confirmed by ablation, and explainability analyses. It suggests that the amplitude-based belief state is the primary decision signal during attacker strategy transitions when the classical belief collapses. The results demonstrate that non-classical uncertainty representations provide better security for IoV networks without additional hardware.