Differentially Private Federated Learning With Importance Client Sampling

As numerous consumer electronics applications like smartphones and wearables generate lots of distributed data daily, consumers desire to safely and efficiently tackle private and isolated data. Federated learning (FL) is hopeful to satisfy the above requirement due to strong data security and applicability to large-scale scenarios. But diverse clients inevitably cause non-independent and identically distributed (non-iid) data among clients, which severely hinders performance analysis. Besides, affected by non-iid data, the participating clients are typically heterogeneous, which induces the client sampling problem. More importantly, albeit FL can enhance privacy via data localization, for highly secret data like physiological data from wearables, FL should possess stronger security to prevent third-party attacks. For data heterogeneity, client sampling, and privacy security, we propose differential privacy (DP) enabled and importance-aware FL algorithm DPFLICS to jointly handle these problems. Specifically, we utilize the truncated concentrated DP to tightly track the end-to-end privacy loss. To attain better sampling, the server selects partial clients with the probability derived from our importance client sampling. Moreover, to further improve performance, we also leverage the adaptive YOGI optimizer on the server side, which is an adaptive gradient method improved from the widely-used ADAM optimization. Finally, the multiple experiments exhibit the effectiveness of our method.