FedAgent: Federated Learning on Non-IID Data via Reinforcement Learning and Knowledge Distillation

Federated learning (FL) allows distributed client devices to jointly train an efficient global model without transmitting local data, thus providing effective privacy preservation. However, federated learning models trained on non independent and identically distributed (non-IID) data suffer from accuracy degradation and slower convergence speeds. To address the above problems, this paper proposes a novel federated learning framework, FedAgent. Specifically, on the server side, an agent is trained based on a dueling network and a double Deep-Q Network (DQN) to intelligently select the participating clients instead of randomly selecting them. Client selection and knowledge distillation are coupled which retains the global knowledge for local training that exists outside of the selected clients’ data distribution from a global perspective to alleviate the client drift phenomenon. Extensive experiments conducted on multiple datasets demonstrate our framework achieves superior performance over state-of-the-art methods. Furthermore, our framework has a significant advantage in communication efficiency, reducing the number of communication rounds by an average of up to 38.2% on the MNIST, 43.8% on the FashionMNIST, and 17.6% on the CIFAR-10, compared to the best performing models.