Communication-Efficient Federated Learning with Adaptive Number of Participants

Sergey Skorik, Vladislav Dorofeev, Gleb Molodtsov, Aram Avetisyan, Dmitry Bylinkin, Daniil Medyakov, Aleksandr Beznosikov

Keywords: Optimization, Adaptive Participation, Client Selection

Abstract: While communication efficiency is a central challenge in Federated Learning (FL), standard protocols typically rely on a fixed, heuristically chosen number of participating clients per round. This rigid approach often leads to redundant communication in easy optimization stages or insufficient aggregation in heterogeneous regimes. In this work, we propose Intelligent Selection of Participants (ISP), an adaptive algorithm that dynamically optimizes the number of active clients to maximize communication efficiency without compromising convergence. Theoretically, we derive a convergence bound for the non-convex setting, revealing that the required number of participants scales with the gradient heterogeneity, rather than the total number of devices in the network. Guided by this insight, ISP speculatively adjusts the participation budget based on real-time training dynamics. ISP achieves consistent communication savings of up to 30\% while matching the final accuracy of full-budget baselines. Furthermore, detailed ablation studies highlight the robustness of our adaptive criterion, establishing the dynamic selection of client count as a critical, distinct optimization task in federated systems.

Primary Area: optimization

Submission Number: 7155