Publication Details
Overview
 
 
, Linh Trung Nguyen, Oliver Y Chen, Jeroen Van Schependom, ,
 

Chapter in Book/ Report/ Conference proceeding

Abstract 

Federated Learning (FL) is a distributed machine learning paradigm that enables multi-client collaborative training without data leaving local devices, thereby effectively protecting data privacy. However, when faced with challenges such as non-independent and identically distribution (Non-IID) data and external attacks (e.g., gradient leakage attacks and Byzantine attacks), traditional FL methods often encounter convergence difficulty and performance degradation. To alleviate these issues, this paper proposes a decentralized and personalized FL method called PDD-FANs, which employs a robust aggregation mechanism based on generative adversarial networks (GANs) to enhance the model's robustness when training with Non-IID data and under malicious attack. To validate the performance of PDD-FANs in Non-IID setting, we performed comparative experiments on two benchmark datasets against baseline methods, and we evaluated the ability of PDD-FANs to resist Byzantine attack through an ablation study. The experimental results show that PDD-FANs not only perform better in Non-IID setting but also better maintain the learning capability when subjected to malicious attack.

Reference