FedTrident: Resilient Road Condition Classification Against Poisoning Attacks in Federated Learning
Sheng Liu, Panos Papadimitratos
This paper presents FedTrident, a defense system for federated learning-based road condition classification that protects against poisoning attacks where malicious vehicles deliberately provide false training data. The system uses three key techniques: detecting compromised models by analyzing their internal neural patterns, removing unreliable vehicles based on their historical behavior, and repairing the global model using machine unlearning. FedTrident successfully maintains accurate road condition predictions even when malicious participants try to sabotage the system, outperforming existing defense methods.
federated learningadversarial attackspoisoning attacksautonomous vehicles