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

Crypto.world

> Federated learning lets hospitals train shared AI models without moving patient data, but it remains open to Sybil attacks (one adversary posing as many fake participants) and poisoning attacks…

**arXiv**: 2602.02629

**Published**: 2026-02-02

**Authors**: Rodrigo Tertulino, Ricardo Almeida, Laercio Alencar

## AI summary

## Overview

**Research area:** Privacy-preserving machine learning for healthcare, specifically Blockchain-based Federated Learning (BFL) combined with Self-Sovereign Identity (SSI) standards — Decentralized Identifiers (DIDs) and Verifiable Credentials (VCs) — for Electronic Health Record (EHR) collaboration.

**Technical level:** Intermediate. The paper assumes familiarity with federated learning aggregation, smart contracts, and public-key identity standards, but the architecture is explained layer-by-layer and the algorithms are given in pseudocode.

**Scope (one sentence):** The paper proposes and evaluates a Trustworthy Blockchain-based Federated Learning (TBFL) framework that gates participation in collaborative EHR model training on cryptographic identity verification rather than behavioral reputation, tested on the MIMIC-IV database with mortality prediction.

## What This Paper Is About

Federated learning lets hospitals train shared AI models without moving patient data, but it remains open to Sybil attacks (one adversary posing as many fake participants) and poisoning attacks (corrupted model updates). Existing blockchain-based federated learning systems mostly defend against these threats *after the fact*, by scoring participants' reputation from the statistical quality of their gradients or their past behavior. The authors argue this is backwards for healthcare: a newly licensed hospital has no reputation, while a patient attacker can build one before striking. Their goal is to flip the trust model from "trust in behavior" to "trust in identity," so that only cryptographically authenticated institutions — for example, ones holding a credential from a Ministry of Health or a medical licensing board — can ever submit a model update.

## Key Contributions

1. **Gap analysis:** The authors systematically identify the absence of a priori cryptographic authentication in current blockchain-based federated learning schemes as a primary vulnerability enabling both Sybil and poisoning attacks, and contrast this with the dominant post-hoc reputation-scoring approach.
2. **A three-layer TBFL architecture:** An Identity Layer built on W3C DIDs and VCs, a Coordination Layer of Ethereum smart contracts acting as lightweight gatekeepers, and a Learning Layer running FedProx off-chain — with trust anchored in institutional credentials issued by a Trusted Issuer.
3. **Two formalized algorithms:** Algorithm 1 specifies on-chain smart-contract access control (a registry mapping worker addresses to authorization status, with `SubmitUpdate` rejecting unauthorized senders); Algorithm 2 specifies the federated orchestration loop, where a local update is transmitted to the aggregation server only if the corresponding blockchain transaction is confirmed.
4. **An empirical evaluation on MIMIC-IV** reporting Sybil-attack neutralization, mortality-prediction performance, computational overhead, and operational cost for the framework.

## Main Findings

- **Sybil attacks fully blocked:** The framework is reported to neutralize 100% of Sybil attacks, because each participant must present a Verifiable Presentation tied to a credential corresponding to exactly one verified institution, checked deterministically by the smart contract before any update is accepted.
- **Predictive performance preserved:** Clinical utility holds up under the identity-gated federation, with AUC = 0.954 and Recall = 0.890 on the mortality prediction task.
- **Negligible computational overhead:** The identity and blockchain layers add less than 0.12% overhead relative to the federated training computation.
- **Low operational cost:** Total operational costs are approximately $18 for 100 training rounds across multiple institutions, achieved by writing only boolean authorization flags and cryptographic hashes/CIDs on-chain rather than model weight matrices.
- **Deterministic rather than probabilistic security:** Because aggregation is strictly coupled to a confirmed on-chain verification transaction, unauthorized nodes are mathematically precluded from influencing the global model, regardless of attack sophistication — a different guarantee class from reputation scoring.
- **Scalability design validated in principle:** Off-chain storage of parameters (IPFS, with only a compact Content Identifier of typically 46 bytes recorded on-chain) keeps gas costs and block-size constraints manageable.

## Methodology in Plain English

The authors simulate a federation of ten hospitals (K = 10) using the MIMIC-IV critical-care database, version 3.1, hosted on a local PostgreSQL instance. They write a SQL-level extraction that yields 546,028 unique hospital admissions, restricted to adults aged 18 or older, using only the first admission per patient to avoid leakage, and excluding records with a missing in-hospital mortality flag. Twenty-five features are extracted, spanning demographics (age, gender, ethnicity, insurance type), clinical context (admission location, first care unit, Charlson Comorbidity Index), and physiological proxies (length of stay, ICD-10 groupings for sepsis and heart failure).

To mimic the uneven data distribution of real hospitals, records are split across the ten clients using a Dirichlet distribution with concentration α = 0.5, producing strongly non-IID class prevalence. Each client first makes a chronological, stratified 80/20 train/test split; only the training folds are rebalanced with SMOTETomek (combining SMOTE synthetic minority oversampling with Tomek Links removal), while validation and test sets stay imbalanced so results reflect the real distribution. Each client trains a small multi-layer perceptron — 25 inputs, a 64-neuron ReLU hidden layer with 0.2 dropout, a 32-neuron ReLU hidden layer, and a sigmoid output — using SGD with learning rate 0.01, momentum 0.9, and weight decay 1e-5. Training runs 100 global rounds of FedProx with proximal term μ = 0.01, 3 local epochs per round, and batch size 32.

The security layer wraps around this loop. A Trusted Issuer (in practice a regulator or licensing body) verifies each institution off-chain and issues a signed Verifiable Credential; the institution's DID and public key are anchored on Ethereum. Before each round, a participant submits a Verifiable Presentation to the FL Task Contract, which checks the issuer's public key in the Registry Contract, validates the signature, confirms the credential is neither expired nor revoked, and checks that disclosed attributes satisfy the task criteria. Only then does the address enter the round's authorized set. The client trains locally, uploads weights to IPFS, and sends the resulting Content Identifier plus a DID-key signature on-chain; the aggregation server accepts the update only if the transaction is mined and confirmed, and then computes a sample-weighted average.

## Why This Matters

Most blockchain-for-federated-learning research treats trust as something earned through good behavior, which is a poor fit for regulated healthcare where the relevant question is whether a node is a licensed institution at all. This paper reframes that question as a cryptographic admission-control problem, and shows the reframing can be done with essentially no measurable cost to model accuracy or training efficiency and at a trivial dollar cost. It also connects two literatures that have largely run in parallel — SSI/DID standards work and federated learning security — and demonstrates a concrete integration pattern in which the blockchain never touches sensitive data or bulky weights.

**Real-world applications:**

- Multi-hospital mortality or sepsis prediction consortia where participating sites must prove licensure and regulatory standing before contributing updates.
- National or regional health information exchanges that already issue institutional credentials and want to reuse them for AI collaboration, in the spirit of the FHIR-compatible SSI access control work the paper cites.
- Clinical research networks that must demonstrate GDPR/HIPAA-compliant participant authorization to ethics boards and auditors, using the on-chain registry as an immutable participation log.
- Consortium model marketplaces or IoMT ecosystems where eligibility is scoped per task — for example, only credential holders certified for oncological data processing may join an oncology training round.

**Industry relevance:** The design choices that matter commercially are the ones about cost and data placement — content-addressed off-chain storage, hashes and boolean flags on-chain, per-round granular authorization, and a credential revocation registry for license expiry or misconduct. That maps onto how health systems actually procure and govern AI: they need an auditable answer to "which institution contributed to this model, under what authorization, and when."

## Future Directions

- **End-to-end deployment and adversarial testing:** The paper reports Sybil neutralization but the abstract-level results do not present a quantitative poisoning-attack experiment against the gated system; empirically stress-testing poisoning, gradient inversion, and colluding-issuer scenarios is the obvious next step.
- **Credential lifecycle and issuer compromise:** What happens when a Trusted Issuer is itself compromised, or when a credential must be revoked mid-training? The architecture includes a revocation registry, but the practical governance and revocation-propagation latency are open problems.
- **Selective disclosure and zero-knowledge proofs in practice:** The paper notes VCs can use zero-knowledge proofs to prove attributes such as licensure without revealing patient volume or financial data; measuring the real cost of that privacy at the smart-contract and client level remains unexplored.
- **Cross-jurisdiction interoperability:** Mapping credentials from different national Ministries of Health or licensing boards into a single task-eligibility policy — and reconciling it with both GDPR and HIPAA — is unresolved, as is scaling beyond the ten-client simulation used here.

## Target Audience

Graduate students and applied researchers working at the intersection of federated learning, blockchain systems, and health informatics; security and privacy engineers at health systems or health-tech vendors evaluating identity architectures for multi-institutional AI; and standards or compliance architects interested in how W3C DID/VC infrastructure can be operationalized for regulated clinical machine learning. Readers looking for deep cryptographic novelty will find the contribution is architectural and empirical rather than a new primitive; readers looking for a deployable pattern for authenticated federation will find the layer decomposition and cost figures directly useful.

*Note on scope: the supplied paper content is truncated partway through Phase 3 of the workflow description, so detailed tables, baseline comparisons, convergence curves, and the full results discussion from Section 5 are not present in the text reviewed. Any metric not stated above — for example precision, F1, per-round loss values, or comparisons against reputation-based baselines — is not reported in the available content.*

## Authors’ abstract

The digitization of healthcare has generated massive volumes of Electronic Health Records (EHRs), offering unprecedented opportunities 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 promising solution that enables collaborative model training without sharing raw patient data. Despite its potential, FL remains…

[The full abstract on arXiv](https://arxiv.org/abs/2602.02629)

[Read the original paper](https://arxiv.org/abs/2602.02629)

---

Canonical: https://crypto.world/paper/trustworthy-blockchain-based-federated-learning-for-electronic-health-2602.02629
