Detecting falsified lane-changing behavior in connected automated vehicles using physics-guided neural networks

As connected automated vehicles become integrated into modern transportation systems, cybersecurity vulnerabilities increasingly threaten traffic operations and safety. Falsified lane-changing (LC) behavior, induced by cyberattacks such as false data injection, can disrupt proper decision-making and degrade system performance. This study proposes a framework for detecting abnormal LC decisions under cyberattack. The framework comprises two modules: (i) LC decision prediction and (ii) detection of falsified behavior. For LC prediction, this study proposes a physics-guided neural network (PGNN) framework. The PGNN framework integrates a game theory-based LC decision model with a long short-term memory (LSTM) network, complementing physics-based interpretability with data-driven learning capability. Falsified behavior is then detected by monitoring cumulative deviations between predicted LC probabilities and the ground-truth behavior. Evaluation using real-world trajectory data demonstrates high recall and strong accuracy under the evaluated false-data-injection and denial-of-service attacks, with notable performance against the latter compared with existing baselines.