Joint User Association and Resource Allocation for Communication-Efficient Hybrid Federated Learning

Federated Learning (FL) is an efficient distributed machine learning method that builds a shared global model by exchanging model parameters instead of local data, thus protecting user privacy. However, due to the heterogeneity and mobility of user devices, FL experiences high communication latency. To balance learning performance and communication latency, we construct a hybrid federated learning system (HFL) integrating FL and Federated Distillation Learning (FD) in a multi-base station scenario. We propose an optimization problem that jointly considers user mode selection, user association, and resource allocation to minimize communication latency per communication round. To solve this problem, we propose a convex optimization-assisted Double Deep Q-Network (CADDQN) algorithm. Simulation results show that the proposed algorithm outperforms state-of-the-art baselines.