DP-CAKA: Defending Local Model Poisoning Attacks Based on Differential Privacy and Complex Acc-based multi-Krum Algorithm in Distributed Federated Learning

Distributed federated learning models are vulnerable to Byzantine failures because malicious nodes can alter the training direction of the global model, making it deviate from normal training and rendering the global model unusable. To address this issue, we propose a defense scheme based on differential privacy and a Complex Accuracy-based multi-Krum Algorithm (DP-CAKA). First, we construct poisoning attacks against local models and conduct attack experiments using three Byzantine-robust federated learning methods. These experiments significantly increase the error rate of the global model, thereby demonstrating the vulnerability of distributed federated learning systems under Byzantine failures. Subsequently, we add noise to the local gradients using differential privacy to enhance privacy during the aggregation process. In addition, we design a Complex ACC-based Multi-Krum Algorithm (CAKA) that selects a few optimal local gradients to participate in the aggregation, thereby mitigating the negative impact of differential privacy on model accuracy. Experimental results demonstrate that DP-CAKA achieves an effective trade-off between privacy and availability, improving global model accuracy by 7.4% over Krum and Trimmed, and 1.5% over Median under poisoning attacks, by 20.7% over Krum and 0.5% over Median under Gaussian-attacks, and by 11.8% over Trimmed mean and 0.19% over Median under Sign-flipping attacks.