Contrastive Reinforcement Learning for Adaptive Client Selection in Federated Learning

Federated Learning (FL) offers a revolutionary paradigm for data collaboration, allowing multiple clients to train a shared model without directly sharing client data. Client selection is crucial for FL as it directly impacts the balance among data quality, model performance, and privacy protection. However, existing client selection strategies still have limitations in adaptability, resulting in significant room for improvement in training efficiency and effectiveness. In this paper, we propose to utilize reinforcement learning (RL) with contrastive auxiliary task for client selection in federated learning. This paper employs contrastive learning to extract discriminative state-action representations. These representations facilitate an adaptive client selection process, accelerating training convergence towards higher accuracy while minimizing latency penalties. Experimental evaluations conclusively demonstrate that our proposed method significantly reduces overall system latency while dramatically accelerating convergence efficiency.