LAFD: Local-Differentially Private and Asynchronous Federated Learning With Direct Feedback Alignment
Kijung Jung, Incheol Baek, Soo-Hyung Kim, Yon Dohn Chung
Abstract
Federated learning is a promising approach for training machine learning models using distributed data from multiple mobile devices. However, privacy concerns arise when sensitive data are used for training. In this paper, we discuss the challenges of applying local differential privacy to federated learning, which is compounded by limited resources of mobile clients and the asynchronicity of federated learning. To address these challenges, we propose a framework called LAFD, that stands for Local-differentially Private and Asynchronous Federated Learning with Direct Feedback Alignment. LAFD consists of two parts: (a) LFL-DFALS: Local differentially private Federated Learning with Direct Feedback Alignment and Layer Sampling, and (b) AFL-LMTGR: Asynchronous Federated Learning with Local Model Training and Gradient Rebalancing. LFL-DFALS effectively reduces the computation and communication costs via direct feedback alignment and layer sampling. AFL-LMTGR handles the problem of stragglers via local model training, and gradient rebalancing. We demonstrate the performance of LFL-DFALS and AFL-LMTGR through experiments.