CAAFL: A Client-Level Adaptive Acceletative Federated Learning Optimization Algorithm for Heterogeneous Data Scenarios

As an emerging distributed machine learning framework, federated learning has attracted growing attention due to its privacy-preserving capabilities. However, in practical applications, the data from different clients participating in federated learning is often heterogeneous, resulting in the "client drift" issue. This issue requires additional communication epochs to ensure the performance of the global model, consequently increasing the cost of federated learning. To address this issue, we propose a novel client-level adaptive accelerated federated learning (CAAFL) optimization algorithm, which dynamically adjusts the local aggregation step size based on the magnitude of the client’s local gradient change. Compared with the existing algorithms, the proposed algorithm addresses the client drift issue at the client level and requires a smaller amount of additional information exchange. Finally, the proposed algorithm is evaluated on real-world federated learning tasks and its superiority over the existing algorithms is demonstrated.