PDGAN: A Novel Poisoning Defense Method in Federated Learning Using Generative Adversarial Network

Federated learning can complete an enormous training task efficiently by inviting participants to train a deep learning model collaboratively, and the user privacy will be well preserved for the users only upload model parameters to the centralized server. However, the attackers can initiate poisoning attacks by uploading malicious updates in federated learning. Therefore, the accuracy of the global model will be impacted significantly after the attack. To address this vulnerability, we propose a novel poisoning defense generative adversarial network (PDGAN) to defend the poising attack. The PDGAN can reconstruct training data from model updates and audit the accuracy for each participant model by using the generated data. Precisely, the participant whose accuracy is lower than a predefined threshold will be identified as an attacker and model parameters of the attacker will be removed from the training procedure in this iteration. Experiments conducted on MNIST and Fashion-MNIST datasets demonstrate that our approach can indeed defend the poisoning attacks in federated learning.