FedSub: Introducing Class-aware Subnetworks Fusion to Enhance Personalized Federated Learning

TL;DR

FedSub tackles non-IID data in personalized federated learning by marrying class-aware prototypes with class-specific subnetworks and server-side selective fusion. Through dynamic per-class prototype clustering and three fusion strategies, it achieves fast convergence and improved personalization in HAR and mobile-health scenarios, while offering privacy considerations and scalability analysis. Extensive static and dynamic experiments show FedSub consistently outperforming state-of-the-art baselines, with Overlapping Components fusion delivering the strongest gains and LRP-based subnetworks reducing communication overhead. The work also outlines practical future directions, including handling large label spaces, privacy enhancements, and decentralized deployments to broaden applicability.

Abstract

Personalized Federated Learning aims at addressing the challenges of non-IID data in collaborative model training. However, existing methods struggle to balance personalization and generalization, often oversimplifying client similarities or relying too heavily on global models. In this paper, we propose FedSub, a novel approach that introduces class-aware model updates based on data prototypes and model subnetworks fusion to enhance personalization. Prototypes serve as compact representations of client data for each class, clustered on the server to capture label-specific similarities among the clients. Meanwhile, model subnetworks encapsulate the most relevant components to process each class and they are then fused on the server based on the identified clusters to generate fine-grained, class-specific, and highly personalized model updates for each client. Experimental results in three real-world scenarios with high data heterogeneity in human activity recognition and mobile health applications demonstrate the effectiveness of FedSub with respect to state-of-the-art methods to achieve fast convergence and high classification performance.

Figures (8)

• Figure 1: Specificity of users' data patterns. Figures a) and b) display the t-SNE representations of user patterns for the same data classes in HHAR and WESAD, respectively, while Figure c) highlights the variation in user clusters across the different labels available in HHAR.
• Figure 2: High-level data flow in a FedSub federated learning round. Starting from the botton, each client $u_i$ sends to the server the prototype $\rho_{u, y}$, subnetwork $\zeta_{u,y}$, and classification score $\omega_{u,y}$ for each local class label $y$. The server updates the prototype clusters and generates personalized model updates for each client by fusing the received subnetworks from each cluster.
• Figure 3: Quality of prototype clustering in terms of DBI and CHI, under varying levels of perturbation.
• Figure 4: Average F1-Score and loss value with 95% confidence intervals in the static scenario.
• Figure 5: Average F1-Score and loss value with 95% confidence intervals in the dynamic scenario.