A decentralized privacy-preserving framework for diabetic retinopathy detection using federated learning and blockchain
Omar Dib
Abstract
Diabetic Retinopathy (DR) detection in distributed telemedicine environments requires secure, scalable, and privacy-preserving solutions. Traditional federated learning (FL) relies on a central server, raising concerns about data privacy and system trust. We propose a novel serverless framework, FL-BC-SMPC-SMOTE, that integrates deep learning, FL, secure multi-party computation (SMPC), the Synthetic Minority Over-sampling Technique (SMOTE), Blockchain (Hyperledger Fabric), and the InterPlanetary File System (IPFS) to address these challenges. Using the APTOS 2019 dataset, we trained CNN-based models (e.g., EfficientNet-B0, ResNet-18) across 2–10 clients, achieving approximately 90% accuracy without raw data sharing. SMPC eliminates the need for a central aggregator by distributing encrypted model updates among clients, enabling privacy-preserving learning. Blockchain ensures auditable and tamper-resistant aggregation, while IPFS significantly reduces communication overhead—from 64 GB to 100 KB per round. Local SMOTE enhances recall for minority classes by 10–15%, promoting equity in DR severity classification. Compared to differentially private baselines (52.18% accuracy), our framework delivers a robust balance of performance, privacy, and fairness. This GDPR/HIPAA-compliant solution offers a practical and trustworthy approach to decentralized DR detection in real-world telemedicine settings.