Fed-RCPA: Representative Class Prototype Aggregation-based One-Shot Federated Learning

Abstract

Federated Learning enables numerous rounds for server-clients communication, leading to critical challenges such as network instability, and vulnerability to privacy attacks, and high communication overhead. One-shot FL (OFL) represents a Federated Learning paradigm that overcomes above drawbacks by a single round of  server-clients communication. However, existing OFL faces expensive computation cost and fails to address non Independent and Identically Distributed (non-IID) data efficiently and robustly. To overcome these issues, this work introduces Fed-RCPA, a novel OFL that capture Representative Class Prototypes for aggregation. Using global feature statistics enables heterogeneity-resistant and training-free aggregation process. In addition, we ameliorate the prototypes quality using  Dual Tree Complex Wavelet Transform (DTCWT) for efficient aggregation.  Furthermore, we incorporate explainability using Grad-CAM tool for improving trust and the transparency. Fed-RCPA application to two different scenarios, using benchmarks and real medical case,  demonstrates the method effectiveness across varying degrees of data heterogeneity settings. Experimental evaluation on CIFAR-10, CIFAR-100, and SVHN demonstrates that Fed-RCPA achieves accuracies of 65.55%, 41.08%, and 61.94%, respectively, while maintaining stable performance across the investigated non-IID heterogeneity settings. In addition, its application to chest X-ray images for pneumonia detection achieves 69.55% test accuracy with 100% sensitivity for pneumonia cases. These results demonstrate the robustness of Fed-RCPA to data heterogeneity and its applicability to both benchmark and real-world medical imaging scenarios.