FortiFed-SP: Client-Side Defense Against Label-Flipping Attacks in Federated Learning

Federated Learning enables collaborative model training across decentralized clients without requiring direct data exchange, thereby preserving data privacy. Despite this advantage, FL remains susceptible to label-flipping attacks, where malicious clients intentionally alter class labels while keeping input features unchanged, leading to compromised global model performance and misleading predictions. Existing defenses, predominantly server-side aggregation strategies, often fail to effectively counter such subtle yet impactful attacks. To address this issue, we propose FortiFed-SP, a client-side defense framework designed to enhance the robustness of local models against label-flipping attacks. FortiFed-SP combines two core components: Meta-Learning for Adaptive Optimization, which improves the adversarial robustness of client updates, and Attack-Resilient Global Knowledge Distillation, which facilitates structure-preserving and noise-tolerant knowledge transfer from the global model, even in the presence of corrupted labels. Extensive experiments on the CIFAR-10 and CIFAR-100 datasets demonstrate that FortiFed-SP significantly improves model robustness, outperforming state-of-the-art defense baselines by up to 7.97 % across various aggregation strategies.