FedFTL-R: Feature-Interactive Federated Transfer Learning from a Reinforcement Learning Perspective

Federated Learning (FL) has garnered significant attention as a promising solution to address privacy concerns in centralized learning (CL). However, traditional FL models often experience excessive communication overhead, and efforts to improve model performance frequently compromise interpretability. Existing methods typically struggle to balance these challenges. This paper introduces the FedFTL-R framework, designed to enhance FL model performance while managing communication costs and ensuring interpretability. In this framework, clients transmit low-dimensional features and labels generated by the Federated Transfer Learning (FTL) submodule. The server reconstructs these features using a reinforcement learning (RL)-driven automatic feature extraction module, resulting in a highly interpretable and traceable feature space. Subsequently, the server efficiently trains a task-specific submodule. Experimental results demonstrate that FedFTL-R significantly reduces communication overhead and improves classification accuracy by up to 3.21% compared to relevant baselines. Furthermore, by incorporating a broader range of operators, FedFTL-R effectively enhances both interpretability and scalability within the feature space.