Clean-Label Poisoning Attacks and Defense Game in Internet of Vehicles Federated Learning
Clean-label Poisoning Attack and Defense Game for Federated Learning in Vehicular Networks
This study targets Basic Safety Message (BSM) anomaly detection in Internet of Vehicles federated learning and proposes a feature-space clean-label poisoning attack: benign samples are perturbed toward the malicious feature extreme, preserving labels while destroying class separability. On the defense side, feature-consistency relabeling is used to detect suspicious samples, and a game-theoretic framework characterizes the attack-defense equilibrium. Results show that the attack scales with the number of controlled nodes, while the defense suppresses the attack success rate without reducing clean accuracy and pushes the equilibrium toward regions unfavorable to the attack.