Federated Learning Under Adversarial Silence Attacks

Federated learning is a decentralized machine learning framework that trains a high-quality model without transmitting local raw data. Because of the large scale of a federated learning system, an emerging line of research is devoted to tackling arbitrary client unavailability. Existing work considers relatively benign unavailability patterns. In this paper, we consider adversarial client unavailability wherein the system adversary can adaptively silence a subset of clients. We refer to this threat model as client silence attacks. We show that simple variants of FedAvg or FedProx, albeit completely agnostic to the silence fraction, converge to a stationary point of both nonconvex and strongly convex global objectives. Conversely, we establish the fundamental limits and show that the convergence rates of the simple variants can achieve those limits. Furthermore, the convergence speeds of the FedAvg or FedProx variants that we show are the best possible for any first order method that only has access to noisy gradients. Our proofs build upon a tight analysis of the selection bias that persists in the entire learning process. We validate our theoretical prediction through numerical experiments on synthetic and real-world data sets.

History: Accepted by Ram Ramesh, Area Editor for Data Science & Machine Learning.

Funding: This work was supported by the National Key Research and Development Program of China [Grant 2024YFA1015800], the Army Research Laboratory [Grant W911NF-23-2-0014], the Division of Computing and Communication Foundations [Grants 2144593, 2340482], and the National Natural Science Foundation of China [Grant 12101353].

Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information (https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2024.1017) as well as from the IJOC GitHub software repository (https://github.com/INFORMSJoC/2024.1017). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/.