A Decentralized Identity Framework for Secure Federated Learning in Healthcare

: Federated learning (FL) enables collaborative model training across decentralized datasets, thus maintaining the privacy of training data. However, FL remains vulnerable to malicious actors, posing significant risks in privacy-sensitive domains like healthcare. Previous machine learning trust frameworks, while promising, often rely on resource-intensive blockchain ledgers, introducing computational overhead and metadata leakage risks. To address these limitations, this study presents a novel Decentralized Identity (DID) framework for mutual authentication that establishes verifiable trust among participants in FL without dependence on centralized authorities or high-cost blockchain ledgers. The proposed system leverages Decentralized Identifiers (DIDs) and Verifiable Credentials (VCs) issued by a designated Regulator agent to create a lightweight, cryptographically secure identity management layer, ensuring that only trusted, authorized participants contribute to model aggregation. The framework is empirically evaluated against a severe model poisoning attack. Without the framework, the attack successfully compromised model reliability, reducing overall accuracy to 0.68 and malignant recall to 0.51. With the proposed DID-based framework, malicious clients were successfully excluded, and performance rebounded to near-baseline levels (accuracy = 0.95, malignant recall = 0.99). Also, this system achieves significantly improved integrity protection with minimal overhead: total trust overhead per FL round measured 75.5 ms, confirming its suitability for low-latency environments. Furthermore, this latency is approximately 39.7 times faster when compared to the consensus step alone of an optimized distributed ledger-based trust system, validating the framework as a scalable and highly efficient solution for next-generation FL security.