Trustworthy Blockchain-based Federated Learning for Electronic Health Records: Securing Participant Identity with Decentralized Identifiers and Verifiable Credentials

Technical Summary: Trustworthy Blockchain-based Federated Learning for Electronic Health Records

Problem Statement
The digitization of healthcare has generated vast repositories of Electronic Health Records (EHRs), offering significant potential for training Artificial Intelligence (AI) models. However, stringent privacy regulations, such as GDPR and HIPAA, have created data silos that prevent centralized training. Federated Learning (FL) has emerged as a solution to enable collaborative model training without sharing raw data. Despite its privacy-preserving nature, FL in healthcare remains vulnerable to critical security threats, specifically poisoning attacks (where malicious actors inject corrupted gradients) and Sybil attacks (where adversaries create multiple fake identities to influence the consensus mechanism).

Current Blockchain-based FL (BFL) approaches attempt to mitigate these risks but predominantly rely on post-hoc probabilistic reputation systems. These systems establish trust reactively based on the statistical quality of submitted gradients or historical behavior. This approach creates a "cold start" problem for legitimate new institutions and fails to prevent sophisticated adversaries who build reputation before launching attacks. Crucially, existing solutions lack a priori cryptographic identity verification, leaving the network open to unauthorized participants masquerading as legitimate healthcare entities.

Methodology and Proposed Architecture
The paper proposes a Trustworthy Blockchain-based Federated Learning (TBFL) framework that shifts the trust model from "trust in behavior" to "trust in identity" by integrating Self-Sovereign Identity (SSI) standards. The architecture consists of three synergistic layers:

1. Identity Layer (SSI Infrastructure):

   • Decentralized Identifiers (DIDs): Each participating entity (hospitals, labs) generates a globally unique DID linked to a public-private key pair, ensuring self-sovereignty without central authority.
   • Verifiable Credentials (VCs): Trusted Issuers (e.g., Ministries of Health, licensing boards) verify real-world institutional credentials off-chain and issue cryptographically signed VCs. These credentials attest to specific attributes (e.g., "Licensed Level-III Trauma Center") and include metadata for expiration and revocation.

2. Coordination Layer (Blockchain Access Control):

   • Smart Contracts: An Ethereum-based smart contract acts as a deterministic gatekeeper. It maintains a registry of authorized DIDs and their authorization status.
   • Verification Protocol: Before a participant can submit a model update, the smart contract verifies their Verifiable Presentation (derived from their VC). It checks the Issuer's signature, credential validity, and revocation status. Only authorized DIDs are added to the whitelist for a specific training round.
   • Efficiency: To avoid high transaction costs, the system does not store model weights on-chain. Instead, local models are stored on IPFS (InterPlanetary File System), and only the Content Identifier (CID) and a digital signature are submitted to the blockchain.

3. Learning Layer (Privacy-Preserving Training):

   • Algorithm: The framework utilizes FedProx, a robust variant of Federated Averaging that adds a proximal term to handle statistical heterogeneity (Non-IID data) common in healthcare.
   • Data Handling: The study uses the MIMIC-IV dataset (546,028 admissions). Data is partitioned across 10 clients using a Dirichlet distribution to simulate Non-IID conditions. A SMOTE-Tomek algorithm is applied locally to address class imbalance in mortality prediction without leaking data.

Key Contributions

1. Identification of the Authentication Gap: The authors systematically analyze current BFL schemes, identifying the reliance on behavioral reputation rather than cryptographic identity as a primary vulnerability enabling Sybil and poisoning attacks.
2. Identity-First TBFL Architecture: The design and implementation of a framework that strictly gates participation on the verification of Verifiable Credentials. This ensures only authenticated, authorized healthcare providers can submit updates.
3. Seamless Integration of DIDs and Smart Contracts: The demonstration of how DIDs can be integrated with Ethereum smart contracts to verify institutional legitimacy before the computationally intensive aggregation process, optimizing both security and efficiency.

Experimental Results
The framework was evaluated over 100 global training rounds using the MIMIC-IV dataset for in-hospital mortality prediction.

• Security Performance: The TBFL framework successfully neutralized 100% of simulated Sybil attacks. In a comparative experiment, an unprotected baseline system suffered a 12.3% drop in accuracy when attacked, whereas the TBFL system maintained stability. The smart contract rejected all unauthorized submission attempts.
• Clinical Utility: The global model achieved robust predictive performance with an AUC of 0.954 and a Recall of 0.890. The model demonstrated a conservative, safety-oriented strategy, effectively identifying high-risk patients (minimizing False Negatives) while maintaining stability across heterogeneous data distributions.
• Operational Overhead: The integration of blockchain verification introduced negligible computational overhead (<0.12%). The verification step added approximately 0.02 seconds per round compared to 17 seconds for local training.
• Economic Feasibility: The total operational cost for 100 training rounds on the Ethereum testnet was approximately $18. The cost structure is strictly linear and invariant to model complexity because only fixed-length hashes are stored on-chain.

Significance and Claims
The paper claims that the TBFL framework provides a secure, scalable, and economically viable ecosystem for inter-institutional health data collaboration. By anchoring trust in cryptographic identity verification rather than behavioral patterns, the framework fundamentally shifts the security paradigm from reactive threat detection to proactive threat prevention.

The authors assert that this approach resolves the tension between data utility and privacy protection by ensuring that only verified entities can contribute, thereby preventing unauthorized nodes from influencing the global model. The study concludes that this architecture offers a robust foundation for collaborative medical intelligence that is compliant with regulatory standards (GDPR, HIPAA) and resilient against sophisticated adversarial attacks, without compromising clinical performance or imposing prohibitive costs.