부산대학교 도서관

Federated learning (FL) enables collaborative machine learning among multiple devices without sharing private data. However, FL systems are vulnerable to poisoning attacks where malicious participants send malicious model updates to compromise the global model's accuracy. To enhance malicious model detection, we propose an EMDG-FL approach that optimizes the threshold used to identify attacks through a Genetic Algorithm (GA). The threshold indicates the degree of divergence between benign and malicious model updates. A tightly tuned threshold improves detection efficiency by reducing false positives and negatives. Our approach also includes a comparison study evaluating EMDG-FL against other defenses from literature across metrics like Accuracy Rate (ACC), Attack Success Rate (ASR) and Loss Rate (LR). Simulation results using two datasets demonstrate that EMDG-FL outperforms prior works in detecting poisoning attacks in FL. The optimized threshold calculation enables more precise and efficient identification of malicious models.