LAFD: Local-differentially Private and Asynchronous Federated Learning with Direct Feedback Alignment

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