FedTrident: Resilient Road Condition Classification Against Poisoning Attacks in Federated Learning

FedTrident defends against targeted label flipping attacks (TLFA) in federated learning road condition classification through neuron-level analysis, GMM clustering filtering, adaptive client scoring, and machine unlearning. Across multiple FL-RCC models, tasks, and configurations, performance is close to the no-attack scenario, exceeding 8 baseline methods by 9.49% and 4.47% on the two most critical metrics respectively, and remains resilient to the proportion of malicious clients, data heterogeneity, and dynamic attacks.