DiabSecure: Federated-TinyML-enabled unified Privacy-Preserving Framework for Diabetes Risk Assessment using Zero-Trust Blockchain

In this paper, we introduce DiabSecure: a decentralized framework based on FL for real-time diabetes risk prediction using wearable healthcare devices. Local TinyML models deployed on each wearable device are trained on-device, with no need to share raw data, ensuring privacy and security. Model updates are aggregated using the popular FedAvg algorithm, achieving a global accuracy of 0.7593 and an F1-score of 0.7691. To further enhance performance, advanced aggregation strategies such as FedSGD and FedOpt are integrated, with FedSGD attaining an accuracy of 0.7563 and F1-score of 0.7692, and FOA achieving the highest accuracy of 0.76 and F1-score of 0.7725. Precision scores for FOA, FedAvg, and FedSGD are 0.774, 0.772, and 0.76 respectively, highlighting robust predictive power. Achieving this level of performance on the challenging dataset [1], which contains imbalanced, high-dimensional, and behaviorally diverse health records—is non-trivial. To ensure secure, tamper-proof updates, blockchain and IPFS manage data integrity and transparent logging. Our system operates entirely on-device using converted TFLite models, supports 6G-ready communication, and maintains continuous learning at the edge. DiabSecure stands out from traditional centralized systems by relying on dynamic, real-time patient data while ensuring privacy. Evaluation confirms the effectiveness of this approach in dynamic healthcare environments, with high precision and F1-scores across all algorithms, making it a promising solution for continuous health monitoring.