FedPA: Generator-Based Heterogeneous Federated Prototype Adversarial Learning

Federated Learning is an emerging distributed algorithm that is designed to collaboratively train the global model without accessing clients' private data. However, heterogeneity of data among clients leads to significant degradation in model performance. Some studies suggest adopting model regularization and using generators to enrich datasets with diverse features can effectively enhance model performance. But current research focuses on regularizing specific modules of the model, failing to achieve regularization across the entire model, and offering limited mitigation of bias from heterogeneous data. Moreover, few methods consider that generators often produce samples with simple features, and the direct use for generating raw data can raise privacy concerns. To solve these challenges, we propose a generator-based heterogeneous Federated Prototype Adversarial Learning framework, named FedPA, which combines prototype learning and lightweight generators to achieve regularization of the entire model. Our generators are designed to generate features rather than raw data, and use prototype learning to find the hard features in an adversarial learning manner, thereby improving model performance. Experimental results show that FedPA improves test accuracy by 3.7% compared to state-of-the-art methods, validating that FedPA can effectively mitigate model bias.