Robust Communication Strategy for Federated Learning by Incorporating Self-Attention

摘要

Federated learning is an emerging machine learning setting, which can train a shared model on large amounts of decentralized data while protecting data privacy. However, the communication cost of federated learning is heavy, especially for mobile devices with higher latency and lower throughput. Although several algorithms have been proposed to reduce the communication cost, they are extremely sensitive to data distribution, even inapplicable to the real client Non-IID data. In this paper, we propose an effective communication strategy for federated learning called FedSAA, which increases the testing performance on Xon-IID data by introducing self-attention mechanism. Two major innovations of our paper are presented here. Firstly, we utilize self-attention mechanism to optimize both the server-to-client and the client-to-client parameter divergence during the model aggregation process so as to improve the model robustness for Non-IID data. Secondly, we adopt the sign compression operator to help data transmission between nodes. The experimental results demonstrate that the model accuracy of our communication-efficient strategy for federated learning with Non-IID data is superior to other communication-efficient algorithms.