FedValidate: A Robust Federated Learning Framework Based on Client-Side Validation
Wenting Zhu, Zhe Liu, Zongyi Chen, Chuan Shi, Xi Zhang, Sanchuan Guo · 2023
Federated Learning (FL), as a distributed machine learning paradigm, has garnered substantial attention in recent years. However, FL does not always provide adequate privacy and robustness guarantees. Due to its distributed nature, adversaries can submit manipulated gradients during training, thereby compromising the integrity and availability of the global model. Most existing defense mechanisms assume that the gradients uploaded by clients are visible to the server, which is incompatible with secure aggregation. To address this challenge, we propose FedValidate, a robust FL framework that evaluates the credibility of target clients based on client-side collaborative validation. It adaptively adjusts the aggregation weights of target clients to enhance the global model's performance. Furthermore, our framework can be seamlessly integrated with relevant privacy-preserving techniques in privacy-enhanced scenarios to prevent gradient leakage, thereby achieving secure FL. Experiments on the MNIST and CIFAR-10 datasets demonstrate that FedValidate effectively resists poisoning attacks.