A Defense Method for Mitigating Poisoning Attacks in Federated Learning Using ACGAN

Federated Learning (FL) has gained significant attention as a novel distributed machine learning paradigm that addresses data privacy concerns. However, its robustness is challenged by the ease with which attackers can launch poisoning attacks by uploading malicious updates to corrupt the global model. To address this issue, we propose a defense method against poisoning attacks in federated learning based on Auxiliary Classifier Generative Adversarial Networks (ACGAN). Firstly, ACGAN is trained on client devices and aggregated using SCAFFOLD on the server. Then, the generators part of ACGAN models are aggregated to generate labeled auxiliary training sets to retrain the discriminator within the ACGAN, which acts as a global classifier model. Finally, the sanitized global model is sent back to the clients. Experiments show that under different attack strengths with a single attacker, our proposed method maintains recall rates above 95 % and overall accuracy above 90 % for the flipped classes on the Fashion-MNIST dataset. In contrast, on the MNIST dataset, both metrics remain above 98%. Under multiple attackers with varying attack strengths, our method maintains recall rates above 75% and overall accuracy above 87 % for the Fashion-MNIST dataset, and above 78% and 95% on the MNIST dataset, respectively.