FedIPSA: Personalized Federated Learning via Importance-Based Parameter Masking and Selective Alignment

Abstract

Federated learning (FL) enables distributed clients to collaboratively train models without sharing raw data, making it well-suited for privacy-sensitive Internet of Things (IoT) environments. However, under heterogeneous [nonindependent and identically distributed (IID)] data distributions, classical FL often suffers from unstable convergence and poor personalization, as repeated global aggregation overwrites client-specific knowledge. Existing personalized FL (PFL) methods mitigate this issue through architectural partitioning or alignment strategies, but typically treat all parameters uniformly, leading to catastrophic forgetting. In this article, we propose FedIPSA, a novel PFL framework that addresses this limitation through fine-grained, parameter-level personalization. FedIPSA integrates two modules: importance-based parameter masking (IPM) and selective alignment mechanism (SAM). IPM estimates parameter importance using synaptic intelligence (SI) to identify critical client-specific parameters, while SAM preserves their local knowledge and selectively aligns the remaining globally initialized parameters with the previous local model. Experiments on CIFAR-10/100, FEMNIST, and Shakespeare demonstrate that FedIPSA outperforms state-of-the-art methods by up to 5.09% in large-scale settings while ensuring superior training stability and convergence. Furthermore, it achieves a 38.73% reduction in per-round execution time compared to recent alignment-based frameworks.