Skip to content

Personalized Federated Learning via Residual Self-Distillation for Heterogeneous Internet of Things Environments

Dec 2026 · IEEE Internet of Things Journal · Vol 13, pp. 58304-58320 · 0 citations · 55 references

Abstract

Personalized federated learning (PFL) has emerged as a promising paradigm for Internet of Things (IoT) applications, enabling client-specific model training under non-IID data and device heterogeneity while preserving data privacy. However, in heterogeneous IoT environments, resource variability, intermittent client availability, and communication heterogeneity often undermine personalization quality and learning stability. To address this issue, we propose personalized federated learning via residual self-distillation (PRSD), a PFL framework based on residual self-distillation for heterogeneous IoT environments. Specifically, PRSD introduces a constraint-aware client selection mechanism that jointly considers resource feasibility, residual-teacher stability, and historical self-distillation gain to better align participation decisions with training dynamics. In addition, a residual-enhanced teacher, constructed by combining the current global model with client-side personalized memory, is used to provide more stable self-distillation guidance for local training. Furthermore, local model updates are decomposed layer-wise into shared low-rank components and residual components with adaptive rank control, which improves communication efficiency while retaining personalized information on the client side. Experimental results on multiple benchmark datasets show that PRSD achieves competitive or superior personalized accuracy, with gains of up to 7.79 percentage points over the best external baseline, while providing a favorable balance between accuracy and communication cost in terms of accumulated uploaded shared parameters.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.