HFL-SkinNet: dermoscopy-based secure multi-class skin disease diagnosis via hybrid semi-supervised FL framework

Skin cancer is a globally prevalent and potentially fatal disease, where early and accurate diagnosis is crucial for improving treatment outcomes. However, traditional centralized deep learning approaches for skin cancer classification pose serious limitations, including risks to patient privacy, reliance on large labeled datasets, and data heterogeneity across healthcare centers. To address these challenges, we propose HFL-SkinNet, a novel Hierarchical Federated Learning-based framework for multi-class skin cancer classification, designed to operate efficiently in decentralized and privacy-sensitive environments. HFL-SkinNet employs a three-layer architecture comprising the IoT layer for secure image acquisition, the Edge layer for local training, and the Cloud layer for federated aggregation. At the edge, a Hierarchical Multi-Branch Convolutional Neural Network (HMB-CNN) is used for robust feature extraction from dermoscopy images. To reduce dependency on labeled data, Semi-Supervised Learning (SSL) with high-confidence pseudo-labeling is employed. A Dual-Stage Attention (DSA) classifier further refines predictions by focusing on salient spatial and channel-wise features. To ensure strong privacy guarantees, the model incorporates differential privacy (ε = 1.0, δ = 1e-5 ) and Paillier homomorphic encryption for secure model updates, while keeping raw patient data local. The framework significantly reduces communication overhead and training latency, supporting scalability across heterogeneous devices. Evaluations on benchmark datasets including ISIC 2018, PH2 and HAM10000 demonstrate that HFL-SkinNet achieves up to 97.6% accuracy, outperforming state-of-the-art FL and non-FL baselines. Overall, HFL-SkinNet offers a privacy-preserving, communication-efficient, and accurate solution for real-world skin cancer diagnosis in federated healthcare systems.