LOGIT: Learning to Generate Gradients for Federated Learning with Arbitrary Client Unavailability

Jinhao Ouyang, Yuan Liu

Keywords: Federated learning, Gradient generation, Client availability

Abstract: Federated learning (FL) enables distributed clients to collaboratively train a shared model while keeping data local. However, in practice, unreliable transmissions, power constraints, and client mobility induce intermittent client unavailability, which biases gradient aggregation and impedes convergence. To address this issue, we propose LOGIT, a gradient-generation framework that learns client-specific gradient trajectories to reconstruct missing updates on the server when clients drop out. Specifically, LOGIT conditions a lightweight generator on each client’s gradient history and the current-round updates from available clients, producing surrogate gradients for unavailable clients and preserving statistical diversity across participants. We further derive a tighter convergence bound and show that LOGIT converges at a rate of $\mathcal{O}(1/\sqrt{T})$, where $T$ is the number of communication rounds. Our experimental results on public datasets validate the effectiveness of LOGIT, demonstrating consistent superiority over baselines, particularly in scenarios with high data heterogeneity and client unavailability.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 9162