Adaptive Local Step Tuning for federated learning under data heterogeneity

Abstract

Federated Learning (FL) enables collaborative model training across decentralized clients while preserving data privacy. However, non-IID data often induces client drift and unstable convergence due to mismatched local training dynamics across clients. Standard FL methods typically utilize fixed local training budgets, leading to computational inefficiency in heterogeneous environments. This paper proposes Adaptive Local Step Tuning (ALST), a lightweight client-autonomous framework that dynamically adjusts local training steps based on an empirical drift proxy estimated from a single probe update. The framework operates entirely on the client side without introducing additional communication overhead or metadata sharing. Experimental results on EMNIST and CIFAR-100 demonstrate that ALST maintains stable convergence while improving the efficiency-performance trade-off under heterogeneous settings. While EMNIST validates the stability of the proposed adaptive mechanism, results on the more challenging CIFAR-100 benchmark show that ALST reduces total local computation by 23.09% while maintaining predictive performance within 0.14% of fixed-work baselines.