Adaptive Model Pruning in Decentralized Federated Learning
Yuchen Mu, Xiaonan Liu, Tharmalingam Ratnarajah · IEEE Transactions on Network Science and Engineering · 2025
Federated Learning (FL) has emerged as a promising decentralized learning framework that facilitates efficient model training across multiple devices, leveraging insights from a variety of data sources while preserving the privacy of devices. However, the distributed architecture of FL introduces high communication cost between the central server and mobile devices, especially for mobile devices far away from the server. Meanwhile, the computation and communication latency can be high on devices with limited computation capability and wireless resources when updating large-scale learning models. Therefore, in this paper, we consider a decentralized distributed FL with model pruning over wireless networks. In each global communication round, the K-means algorithm is deployed to perform dynamic clustering of devices based on their geographical proximity and device-to-device (D2D) communication technique is used to share the updated model weights. Then, we mathematically analyze the computation and communication latency and convergence of the proposed system. To improve convergence rate, guarantee learning accuracy, and decrease computation and communication latency, Karush-Kuhn-Tucker (KKT) conditions are employed to jointly optimize the pruning ratio and bandwidth allocation. Simulation results demonstrate that the proposed FL algorithm necessitates only 88% of the training time required by the system employing equal pruning and 80% of the training time needed by the system without pruning, while achieving a testing accuracy comparable to that of the latter.