HSSF-ABIDE: A Semi-Supervised Federated Framework with Heterogeneous Clients for Multi-Site fMRI

In this work, we adapt and extend the Heterogeneous Semi-Supervised Federated Learning (HSSF) paradigm to a multi-site autism spectrum disorder (ASD) classification using ABIDE data. Each client trains locally with scarce labels and abundant unlabeled samples, while a server aggregates knowledge via logit-based distillation on a small public set. The method couples supervised cross-entropy with temperature-scaled distillation and an unsupervised consistency objective gated by a local confidence threshold. We split the ABIDE dataset into three scanner-defined clients to reflect real-world acquisition heterogeneity. Under a federated learning setting, each client trains a distinct local backbone (ResNet-18, EfficientNet, or MobileNetV2), enabling heterogeneous client models while preserving privacy. Our approach reaches a global accuracy of 0.96 after 300 communication rounds-close to centralized training (0.98) while preserving data locality. At the client level, consistent gains were observed over isolated training indicating effective cross-site knowledge transfer despite model and data heterogeneity. We discuss communication costs, sensitivity to the pseudo-label threshold, and fairness across clients, and we outline directions for personalized federated learning under multimodal extensions.