Decentralized Federated Learning under Resource Constraints: An Empirical Comparison

Abstract

Decentralized federated learning paradigms are increasingly adopted to support scalable and privacy-preserving training across distributed edge systems. However, existing approaches exhibit significant performance variability under realistic system conditions, including statistical heterogeneity, resource constraints, and client churn. In particular, the tradeoffs between centralized federated learning and fully decentralized optimization methods remain insufficiently understood from a system-level perspective. This paper presents a controlled empirical comparison of centralized federated and decentralized learning approaches under heterogeneous and dynamic environments. The evaluated methods are tested across varying degrees of non-independent and identically distributed (non-IID) data, client availability, communication configurations, and resource-constrained device settings to investigate how communication topology, update synchronization, and node participation affect convergence and predictive performance. The results show that decentralized approaches can provide competitive performance under moderate operating conditions, but their effectiveness varies substantially as data heterogeneity and client churn increase, with gossip-based and communication-compressed methods exhibiting additional challenges on constrained edge devices. These findings demonstrate that the suitability of decentralized learning depends strongly on system conditions and highlight the importance of jointly considering statistical heterogeneity, communication behavior, and device availability when designing distributed learning solutions for edge environments.