Trustworthy Blockchain-based Federated Learning for Electronic Health Records: Identity-Gated Participation against Sybil Attacks

arXiv:2602.02629 · cs.CR, cs.AI, cs.LG · Submitted 2026-02-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "Trustworthy Blockchain-based Federated Learning for Electronic Health Records".

Elias: Here's the gist of the paper. The paper proposes a Trustworthy Blockchain-based Federated Learning (TBFL) framework that integrates Self-Sovereign Identity standards to secure participant identity in Electronic Health Record training, which matters because it addresses critical trust and security gaps in collaborative healthcare AI by anchoring trust in cryptographic identity verification rather than probabilistic reputation systems Problem Statement and Motivation 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 vulnerable to poisoning and Sybil attacks, in which malicious participants corrupt the global model or infiltrate the network using fake identities While recent approaches integrate Blockchain technology for auditability, they predominantly rely on probabilistic reputation systems rather than robust cryptographic identity verification Proposed Architecture Layers The TBFL architecture is structured around three synergistic logical layers: the Identity Layer, which implements Self-Sovereign Identity standards; the Coordination Layer, which leverages Blockchain technology for consensus and access control; and the Learning Layer, which executes the federated machine learning protocol The Identity Layer serves as the foundational mechanism for establishing trust, cryptographically verifying the legitimacy of all network participants before any computational tasks commence This layer implements two core standards defined by the World Wide Web Consortium (W3C), specifically designed to enable decentralized, privacy-preserving identity management Decentralized Identifiers (DIDs): Each participating entity in the TBFL network, including hospitals, research institutions, clinical laboratories, and aggregation nodes, generates a globally unique Decentralized Identifier (DID) Verifiable Credentials (VCs): While DIDs establish cryptographic identity, they alone cannot distinguish between a legitimate healthcare institution and a malicious actor The Trusted Issuer is responsible for conducting rigorous off-chain verification of an entity’s real-world credentials, including institutional licenses, regulatory compliance certifications, and ethical review approvals Workflow and Security Enforcement The TBFL workflow follows a carefully designed four-phase security protocol Phase 1 involves Onboarding and Credential Issuance by a Trusted Issuer, where the Issuer creates a Verifiable Credential that formally attests to specific attributes of the healthcare institution Phase 2 is Authentication and Task Selection, where a participant submits a Verifiable Presentation derived from its VC to join the task, triggering deterministic verification by the smart contract Phase 3 is Authenticated Model Update Submission, where model parameters are uploaded to IPFS and signed with the DID private key before submission Phase 4 is Secure Aggregation, where the aggregator performs a final verification of the signature and credential status before incorporating any update into the global model This sequential approach ensures that security is enforced at every stage of the collaborative learning process Empirical Validation and Results The experimental results demonstrate that integrating SSI-based identity verification enhances security without compromising learning efficiency or clinical utility The TBFL framework successfully neutralized 100% of simulated Sybil attack attempts, wherein malicious actors attempted to register multiple fictitious institutional identities to gain disproportionate influence over model aggregation In comparative experiments with an unprotected baseline system, the introduction of identity verification prevented significant performance degradation that would otherwise occur under adversarial conditions The global model achieved an Area Under the Receiver Operating Characteristic Curve of 0.954, indicating excellent discriminative ability between high-risk and low-risk patients The recall (sensitivity) metric of 0.890 demonstrates that the model successfully identifies approximately 89% of patients who ultimately experience in-hospital mortality The total operational cost for a complete 100-round training cycle amounted to approximately 18 when deployed on the Ethereum testnet, a negligible sum relative to typical healthcare AI research budgets The proposed TBFL framework represents not merely an incremental improvement but rather a fundamental reconceptualization of how trust is established, verified, and enforced in decentralized machine learning networks The research makes several distinct contributions to the intersection of secure machine learning, distributed systems, and healthcare informatics Future research should investigate multi-layered defense mechanisms that complement identity verification with behavioral anomaly detection Specifically, local differential privacy could be applied wherein each institution adds calibrated Gaussian or Laplacian noise to their gradient updates before submission Transition to Permissioned Distributed Ledger Technologies such as Hyperledger Fabric offers potential advantages for real-world healthcare consortia by providing fine-grained, role-based access control and optimized consensus mechanisms Extension to Multi-Modal Clinical Data would significantly expand its clinical utility but introduces substantial technical challenges regarding heterogeneous aggregation problems The path forward requires continued interdisciplinary collaboration among computer scientists, healthcare professionals, legal scholars and ethicists and policymakers creating governance structures that enable beneficial innovation while protecting vulnerable populations The experimental results presented in Section 5 validate the TBFL framework across three fundamental pillars that any healthcare-appropriate collaborative learning system must satisfy Security and Attack Resistance.

Nadia: The TBFL architecture successfully neutralized 100% of simulated Sybil attack attempts, wherein malicious actors attempted to register multiple fictitious institutional identities to gain disproportionate influence over model aggregation In comparative experiments with an unprotected baseline system, the introduction of identity verification prevented significant performance degradation that would otherwise occur under adversarial conditions Statistical analysis confirmed this security benefit with high confidence (t = 17.29, p < 0.001, Cohen’s d = 2.87 indicating a large effect size) Clinical Utility and Patient Safety.

Elias: Beyond mere algorithmic stability and security, the global model achieved clinically meaningful predictive performance that would be suitable for real-world clinical decision support applications Specifically, the model attained an Area Under the Receiver Operating Characteristic Curve of 0.954, indicating excellent discriminative ability between high-risk and low-risk patients The recall (sensitivity) metric of 0.890 demonstrates that the model successfully identifies approximately 89% of patients who ultimately experience in-hospital mortality Operational Viability and Economic Sustainability.

Nadia: Addressing a persistent criticism of blockchain-based systems, namely, that they introduce prohibitive computational overhead and unsustainable operational costs, our comprehensive analysis demonstrated that identity verification introduces negligible additional latency Furthermore, the economic analysis revealed a strictly linear, highly predictable cost structure The total operational expense for a complete 100-round training cycle amounted to approximately 18 when deployed on the Ethereum testnet, a negligible sum relative to typical healthcare AI research budgets that often exceed hundreds of thousands of dollars Theoretical and Practical Contributions From a theoretical perspective, we formalized the security properties of identity-verified federated learning through two complementary algorithms (Algorithm 1 for on-chain access control and Algorithm 2 for off-chain learning orchestration) that establish deterministic, mathematically provable security guarantees rather than probabilistic defenses By explicitly coupling the model aggregation step with blockchain transaction confirmation, such that updates are incorporated into the global model if and only if the smart contract successfully verifies the submitter’s credentials, we mathematically preclude the possibility of unauthorized nodes influencing the global model From a practical standpoint, we demonstrated that W3C-standardized SelfSovereign Identity technologies, specifically Decentralized Identifiers (DIDs) and Verifiable Credentials (VCs), can be seamlessly integrated into real-world machine learning workflows without requiring fundamental architectural changes to existing federated learning protocols or frameworks This backward compatibility is crucial for adoption, as it enables healthcare institutions to enhance the security posture of their collaborative learning initiatives without abandoning existing technical infrastructure Furthermore, our hybrid architectural design provides a reusable blueprint for scalable blockchain-based machine learning systems across diverse application domains Limitations and Future Research Our current security evaluation focused primarily on external threats, specifically Sybil attacks originating from unauthorized entities attempting to infiltrate the federation and basic poisoning attacks in which adversaries inject obviously corrupted gradient updates However, real-world adversaries may employ substantially more sophisticated attack strategies Of particular concern are insider threats, wherein authorized but malicious participants deliberately submit poisoned model updates designed to introduce subtle biases or backdoor triggers Future research should investigate multi-layered defense mechanisms that complement identity verification with behavioral anomaly detection Specifically, longitudinal reputation systems could be developed wherein institutions accumulate trust scores based on the consistency and quality of their historical contributions Alternatively, Zero-Knowledge Proofs (ZKPs) or Secure Multi-Party Computation (SMPC) could enable participants to prove that their model updates satisfy certain quality constraints without revealing the actual gradient values Transition to Permissioned Distributed Ledger Technologies such as Hyperledger Fabric offers potential advantages in these dimensions by providing fine-grained, role-based access control and optimized consensus mechanisms Extension to Multi-Modal Clinical Data would significantly expand its clinical utility but introduces substantial technical challenges regarding heterogeneous aggregation problems Dynamic Credential Management and Temporal Trust Evolution.

Paper summary: Nadia: So we're looking at this paper, "Trustworthy Blockchain-based Federated Learning for Electronic Health Records: Identity-Gated Participation against Sybil Attacks." Basically, the main idea is to use Self-Sovereign Identity standards to secure who can join a collaborative training network for AI models on Electronic Health Records.

Elias: That's right. The thesis is that it tackles the big problem in federated learning for healthcare, where malicious participants can sneak in using fake identities, which we call Sybil attacks, or they can poison the model and ruin everything.

Nadia: And what this paper claims is that instead of relying on those probabilistic reputation systems that most current blockchain approaches use, this framework anchors trust in cryptographic identity verification.

Elias: Exactly. It proposes a Trustworthy Blockchain-based Federated Learning framework, or TBFL, which uses decentralized identifiers and verifiable credentials to establish identity before any training even starts.

Priya: But I want to know what that means for the actual data. If you're verifying an identity cryptographically, how does that actually translate into reliable health data?

Nadia: Well, the paper lays out this architecture with three layers: an Identity Layer using Self-Sovereign Identity standards, a Coordination Layer on the Blockchain for consensus and access control, and then the Learning Layer where the actual federated machine learning protocol runs.

Elias: The Identity Layer is key there. It creates Decentralized Identifiers or DIDs for every entity in this network, like hospitals or research labs. But just having a DID isn't enough on its own; they also use Verifiable Credentials, or VCs, which act like official digital certificates.

Priya: So these credentials are what verify the real-world thing—like making sure the hospital is actually licensed or has the proper regulatory approvals. What's the mechanism for that verification?

Nadia: The paper describes a four-phase security protocol for this workflow. First, there's onboarding where a Trusted Issuer creates a Verifiable Credential attesting to specific attributes of an institution. Then, there’s authentication and task selection, where the participant submits their Verifiable Presentation for deterministic verification by the smart contract.

Elias: Then comes authenticated model update submission, where those model parameters are uploaded and signed with the DID private key before they even get into the system. And finally, secure aggregation happens where there's a final verification of that signature and credential status before anything gets added to the global model.

Paper summary: Priya: That sounds like a lot of steps for something that needs to be fast in a learning process. Are these cryptographic checks going to slow down how quickly the AI models learn from the data?

Nadia: The results suggest that integrating this SSI-based identity verification actually enhances security without compromising learning efficiency or clinical utility. They found that it successfully neutralized one hundred percent of simulated Sybil attack attempts where malicious actors tried to register multiple fake institutional identities <ref:2602.02629#pg2>.

Elias: And they compared it against an unprotected baseline system, and the introduction of identity verification prevented significant performance degradation that would otherwise happen when there were adversarial conditions in the learning process.

Priya: So, beyond just stopping attacks, how good is this resulting global model actually being for real patient care? What are the metrics showing here?

Nadia: The global model achieved an Area Under the Receiver Operating Characteristic Curve of zero point nine five four <ref:2602.02629#pg1>. That number indicates excellent discriminative ability between high-risk and low-risk patients, which is clinically meaningful.

Elias: And they also found that the recall metric, which is sensitivity, was zero point eight nine zero, meaning the model successfully identifies about eighty-nine percent of patients who ultimately experience in-hospital mortality. That's a solid number for predictive performance in this context.

Priya: It’s interesting that the operational cost analysis also came out relatively low when they ran the simulation on the Ethereum testnet. What was their estimate there?

Nadia: They found that the total operational cost for a complete one hundred training round cycle amounted to approximately when deployed on the Ethereum testnet, which is pretty negligible compared to typical healthcare AI research budgets <ref:2602.02629#pg2>.

Elias: So, they’re claiming this isn't just a theoretical security fix; it has practical implications for how we build and deploy these collaborative learning systems in a real setting.

Nadia: It does. This paper focuses on how trust is established through verifiable institutional identity rather than relying on statistical reputation metrics or post-hoc behavioral analysis, which the authors point out is where current blockchain approaches fall short, especially in critical healthcare environments.

Paper summary: Elias: The title of this paper, "Trustworthy Blockchain-based Federated Learning for Electronic Health Records: Identity-Gated Participation against Sybil Attacks," really captures the core contribution of anchoring trust in verifiable credentials within a federated learning structure.

Priya: If we look at what this means practically for institutions, it suggests a way to pool clinical intelligence across different hospitals while respecting those strict patient privacy regulations like GDPR and HIPAA.

Nadia: That's right. It gives hospitals and research institutions a foundation for collaborative learning that is more robust because they are verifying the legitimacy of their partners cryptographically before they ever share model updates or patient data insights.

Elias: The implication here is moving away from trust being established reactively through statistical quality of submitted gradients, toward proactive authentication based on verifiable credentials.

Priya: But I do have to ask what the authors flagged as a limitation. What's the gap they admitted they didn't fully cover with this specific framework?

Nadia: They did admit that their current security evaluation focused primarily on external threats, specifically Sybil attacks from unauthorized entities trying to infiltrate the federation.

Elias: And they also mentioned that while it’s deterministic against identity attacks, real-world adversaries might employ more sophisticated attack strategies, particularly insider threats where authorized but malicious participants deliberately submit poisoned model updates.

Nadia: So, to wrap up what we've heard about the paper "Trustworthy Blockchain-based Federated Learning for Electronic Health Records: Identity-Gated Participation against Sybil Attacks," this framework fundamentally rethinks how we build trust in decentralized machine learning networks.

Elias: The paper argues that by integrating Self-Sovereign Identity standards, specifically Decentralized Identifiers and Verifiable Credentials, you can establish deterministic security guarantees instead of relying on probabilistic defenses.

Priya: It shows that this approach doesn't just stop the attacks; it also results in a high-performing global model with a recall of zero point eight nine zero, showing good predictive capability for patient risk identification.

Nadia: The key point is that we can anchor trust in cryptographically verifiable institutional identity instead of purely statistical reputation metrics or post-hoc behavioral analysis.

Elias: The implication for the world is providing a reusable blueprint for scalable blockchain-based machine learning systems across different domains, like healthcare, by making the security posture more transparent and legally defensible.

Conclusion: Nadia: So we’re wrapping up this look at the paper, "Trustworthy Blockchain-based Federated Learning for Electronic Health Records: Identity-Gated Participation against Sybil Attacks." Basically, this work is about building a system where you can train AI models using patient health data across different institutions without letting fake people sneak into the network.

Elias: Right. The core idea here is moving away from just guessing if a participant is honest based on their past behavior and instead using cryptographic proof of who they are.

Nadia: It’s about anchoring that trust in verifiable identity standards, specifically things like Decentralized Identifiers and Verifiable Credentials, instead of relying on those probabilistic reputation systems that many current blockchain solutions use.

Elias: Exactly. They’re proposing a framework where you prove your credentials before you even start the training process.

Nadia: The paper shows how this works across three layers—identity, coordination on the blockchain, and then the actual learning part—making sure security is checked at every single step of that whole workflow.

Elias: And what they found is that this approach doesn't just stop those obvious Sybil attacks where people try to register fake institutions; it actually keeps the model performing well when there are bad things happening in the network.

Nadia: They got a very high recall metric, about eighty-nine percent, for identifying patients who might end up with in-hospital mortality. That’s a solid number for clinical utility.

Elias: And I looked at their cost analysis—they showed the operational expense for a hundred training rounds on the Ethereum testnet was really low, around eighteen units. That addresses a big concern about blockchain systems having too much overhead.

Nadia: It seems like they’ve found a way to get strong security and good clinical results without making things prohibitively expensive or slow for actual medical research.

Elias: It fundamentally rethinks how we build trust in these decentralized machine learning networks by using verifiable credentials instead of just statistical reputation metrics.

Nadia: And that leads us to thinking about what comes next, especially since the authors did mention where their testing stops—specifically, they focused mostly on external attacks and haven't fully addressed those insider threats from authorized but malicious participants.

Rodrigo Tertulino, Ricardo Almeida, Laercio Alencar

Federal Institute of Education, Science, and Technology of Rio Grande do Norte (IFRN)

cs.CR, cs.AI, cs.LG

Submitted: 2026-02-02

Updated: 2026-10-31

Comments: 28 pages, 10 figures, 14 tables. v2: substantially revised. Fixes credential replay after revocation in the registry contract, adds testnet deployments (Ethereum Sepolia, Arbitrum Sepolia), credentialed-insider and label-flipping experiments, and replaces the illustrative results of v1 with multi-seed results. Code and run logs: https://doi.org/10.5281/zenodo.23836753

Code: https://github.com/rodrigoronner/TBFL-EHR-Framework

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: The paper proposes a Trustworthy Blockchain-based Federated Learning (TBFL) framework that integrates Self-Sovereign Identity standards to secure participant identity in Electronic Health Record…

Key concepts

Self-Sovereign Identity (SSI)
SSI allows individuals or entities to control their own digital identities using decentralized identifiers (DIDs) and verifiable credentials (VCs). This means participants in the learning network can cryptographically prove who they are, rather than relying on a central authority to manage their credentials, enhancing privacy and trust.
Decentralized Identifiers (DIDs)
DIDs are globally unique identifiers that participants generate for themselves. They act as a decentralized way to establish cryptographic identity in the TBFL network. These DIDs form the basis for verifying the legitimacy of hospitals and researchers before they can participate in model training tasks.
Verifiable Credentials (VCs)
VCs are digital documents issued by a trusted entity, like a hospital or research institution, that attest to specific attributes of that entity. These credentials formally verify real-world credentials such as institutional licenses and regulatory compliance, providing the necessary proof for identity verification in the learning process.

Terminology

Summary

The paper proposes a Trustworthy Blockchain-based Federated Learning (TBFL) framework that integrates Self-Sovereign Identity standards to secure participant identity in Electronic Health Record training, which matters because it addresses critical trust and security gaps in collaborative healthcare AI by anchoring trust in cryptographic identity verification rather than probabilistic reputation systems

Problem Statement and Motivation

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 vulnerable to poisoning and Sybil attacks, in which malicious participants corrupt the global model or infiltrate the network using fake identities While recent approaches integrate Blockchain technology for auditability, they predominantly rely on probabilistic reputation systems rather than robust cryptographic identity verification

Proposed Architecture Layers

The TBFL architecture is structured around three synergistic logical layers: the Identity Layer, which implements Self-Sovereign Identity standards; the Coordination Layer, which leverages Blockchain technology for consensus and access control; and the Learning Layer, which executes the federated machine learning protocol The Identity Layer serves as the foundational mechanism for establishing trust, cryptographically verifying the legitimacy of all network participants before any computational tasks commence This layer implements two core standards defined by the World Wide Web Consortium (W3C), specifically designed to enable decentralized, privacy-preserving identity management Decentralized Identifiers (DIDs): Each participating entity in the TBFL network, including hospitals, research institutions, clinical laboratories, and aggregation nodes, generates a globally unique Decentralized Identifier (DID) Verifiable Credentials (VCs): While DIDs establish cryptographic identity, they alone cannot distinguish between a legitimate healthcare institution and a malicious actor The Trusted Issuer is responsible for conducting rigorous off-chain verification of an entity’s real-world credentials, including institutional licenses, regulatory compliance certifications, and ethical review approvals

Workflow and Security Enforcement

The TBFL workflow follows a carefully designed four-phase security protocol Phase 1 involves Onboarding and Credential Issuance by a Trusted Issuer, where the Issuer creates a Verifiable Credential that formally attests to specific attributes of the healthcare institution Phase 2 is Authentication and Task Selection, where a participant submits a Verifiable Presentation derived from its VC to join the task, triggering deterministic verification by the smart contract Phase 3 is Authenticated Model Update Submission, where model parameters are uploaded to IPFS and signed with the DID private key before submission Phase 4 is Secure Aggregation, where the aggregator performs a final verification of the signature and credential status before incorporating any update into the global model This sequential approach ensures that security is enforced at every stage of the collaborative learning process

Empirical Validation and Results

The experimental results demonstrate that integrating SSI-based identity verification enhances security without compromising learning efficiency or clinical utility The TBFL framework successfully neutralized 100% of simulated Sybil attack attempts, wherein malicious actors attempted to register multiple fictitious institutional identities to gain disproportionate influence over model aggregation In comparative experiments with an unprotected baseline system, the introduction of identity verification prevented significant performance degradation that would otherwise occur under adversarial conditions The global model achieved an Area Under the Receiver Operating Characteristic Curve of 0.954, indicating excellent discriminative ability between high-risk and low-risk patients The recall (sensitivity) metric of 0. p. 0

Where the paper says it

Our results show the framework successfully neutralizes 100% of Sybil attacks, achieves robust predictive performance (AUC = 0. 954, Recall = 0.

p. 0
Where the paper says it

Our results show the framework successfully neutralizes 100% of Sybil attacks, achieves robust predictive performance (AUC = 0. 954, Recall = 0.

890 demonstrates that the model successfully identifies approximately 89% of patients who ultimately experience in-hospital mortality The total operational cost for a complete 100-round training cycle amounted to approximately 18 when deployed on the Ethereum testnet, a negligible sum relative to typical healthcare AI research budgets

Conclusion and Future Directions

The proposed TBFL framework represents not merely an incremental improvement but rather a fundamental reconceptualization of how trust is established, verified, and enforced in decentralized machine learning networks The research makes several distinct contributions to the intersection of secure machine learning, distributed systems, and healthcare informatics Future research should investigate multi-layered defense mechanisms that complement identity verification with behavioral anomaly detection Specifically, local differential privacy could be applied wherein each institution adds calibrated Gaussian or Laplacian noise to their gradient updates before submission Transition to Permissioned Distributed Ledger Technologies such as Hyperledger Fabric offers potential advantages for real-world healthcare consortia by providing fine-grained, role-based access control and optimized consensus mechanisms Extension to Multi-Modal Clinical Data would significantly expand its clinical utility but introduces substantial technical challenges regarding heterogeneous aggregation problems The path forward requires continued interdisciplinary collaboration among computer scientists, healthcare professionals, legal scholars and ethicists and policymakers creating governance structures that enable beneficial innovation while protecting vulnerable populations

Summary of Principal Findings

The experimental results presented in Section 5 validate the TBFL framework across three fundamental pillars that any healthcare-appropriate collaborative learning system must satisfy Security and Attack Resistance. The TBFL architecture successfully neutralized 100% of simulated Sybil attack attempts, wherein malicious actors attempted to register multiple fictitious institutional identities to gain disproportionate influence over model aggregation In comparative experiments with an unprotected baseline system, the introduction of identity verification prevented significant performance degradation that would otherwise occur under adversarial conditions Statistical analysis confirmed this security benefit with high confidence (t = 17. p. 1

Where the paper says it

Recent systematic reviews have identified critical vulnerabilities in traditional FL architectures, particularly concerning the verification and authentication of participating entities (Qammar et al. Among the most severe threats are poisoning attacks, wherein malicious actors inject corrupted or adversarially crafted gradients to compromise the integrity of the global model, and Sybil attacks, wherein adversaries create multiple fake identities to gain disproportionate influence over the consensus mechanism and model aggregation process (Samy and Girdzijauskas 2023; Lian et al.

29, p < 0.001, Cohen’s d = 2.87 indicating a large effect size) Clinical Utility and Patient Safety. p. 2
Where the paper says it

In these systems, trust is typically established reactively, based on the statistical quality of submitted gradients or participants’ historical behavior, rather than through proactive institutional authentication. This approach proves insufficient for critical healthcare environments, where verifying the legitimacy of a participating node, for instance, confirming that it represents a licensed hospital or authorized research institution, is paramount to ensuring both data security and clinical safety (Zhan et al.

Beyond mere algorithmic stability and security, the global model achieved clinically meaningful predictive performance that would be suitable for real-world clinical decision support applications Specifically, the model attained an Area Under the Receiver Operating Characteristic Curve of 0.954, indicating excellent discriminative ability between high-risk and low-risk patients The recall (sensitivity) metric of 0.890 demonstrates that the model successfully identifies approximately 89% of patients who ultimately experience in-hospital mortality Operational Viability and Economic Sustainability. Addressing a persistent criticism of blockchain-based systems, namely, that they introduce prohibitive computational overhead and unsustainable operational costs, our comprehensive analysis demonstrated that identity verification introduces negligible additional latency Furthermore, the economic analysis revealed a strictly linear, highly predictable cost structure The total operational expense for a complete 100-round training cycle amounted to approximately 18 when deployed on the Ethereum testnet, a negligible sum relative to typical healthcare AI research budgets that often exceed hundreds of thousands of dollars

Theoretical and Practical Contributions

From a theoretical perspective, we formalized the security properties of identity-verified federated learning through two complementary algorithms (Algorithm 1 for on-chain access control and Algorithm 2 for off-chain learning orchestration) that establish deterministic, mathematically provable security guarantees rather than probabilistic defenses By explicitly coupling the model aggregation step with blockchain transaction confirmation, such that updates are incorporated into the global model if and only if the smart contract successfully verifies the submitter’s credentials, we mathematically preclude the possibility of unauthorized nodes influencing the global model From a practical standpoint, we demonstrated that W3C-standardized SelfSovereign Identity technologies, specifically Decentralized Identifiers (DIDs) and Verifiable Credentials (VCs), can be seamlessly integrated into real-world machine learning workflows without requiring fundamental architectural changes to existing federated learning protocols or frameworks This backward compatibility is crucial for adoption, as it enables healthcare institutions to enhance the security posture of their collaborative learning initiatives without abandoning existing technical infrastructure Furthermore, our hybrid architectural design provides a reusable blueprint for scalable blockchain-based machine learning systems across diverse application domains

Limitations and Future Research

Our current security evaluation focused primarily on external threats, specifically Sybil attacks originating from unauthorized entities attempting to infiltrate the federation and basic poisoning attacks in which adversaries inject obviously corrupted gradient updates However, real-world adversaries may employ substantially more sophisticated attack strategies Of particular concern are insider threats, wherein authorized but malicious participants deliberately submit poisoned model updates designed to introduce subtle biases or backdoor triggers Future research should investigate multi-layered defense mechanisms that complement identity verification with behavioral anomaly detection Specifically, longitudinal reputation systems could be developed wherein institutions accumulate trust scores based on the consistency and quality of their historical contributions Alternatively, Zero-Knowledge Proofs (ZKPs) or Secure Multi-Party Computation (SMPC) could enable participants to prove that their model updates satisfy certain quality constraints without revealing the actual gradient values Transition to Permissioned Distributed Ledger Technologies such as Hyperledger Fabric offers potential advantages in these dimensions by providing fine-grained, role-based access control and optimized consensus mechanisms Extension to Multi-Modal Clinical Data would significantly expand its clinical utility but introduces substantial technical challenges regarding heterogeneous aggregation problems Dynamic Credential Management and Temporal Trust Evolution. Our current implementation treats credentials as valid or revoked, whereas a newer institution may gradually build a valid or revoked credential Future work could explore incorporating temporal trust decay mechanisms wherein credentials gradually lose validity unless periodically renewed through re-certification processes Scope of Defense. It is important to distinguish between identity attacks (Sybil) and insider threats (poisoning by authorized clients) The proposed TBFL architecture provides deterministic protection against the former While FedProx offers partial robustness against data heterogeneity, specific defenses against authorized malicious insiders require complementary robust aggregation layers like Krum or Trimmed Mean

Broader Impact and Concluding Remarks

By demonstrating a viable, empirically validated path toward secure inter-institutional data collaboration, the TBFL framework directly addresses one of the most fundamental barriers to realizing the transformative promise of artificial intelligence in healthcare Enabling hospitals, clinics, research institutions, and public health agencies to pool their collective clinical intelligence while simultaneously respecting patient privacy could catalyze substantial advances across multiple dimensions of healthcare delivery This research represents a fundamental reconceptualization of how trust is established, verified, and enforced in decentralized machine learning networks By anchoring trust in cryptographically verifiable institutional identity rather than purely statistical reputation metrics or post-hoc behavioral analysis, we create a more robust, transparent, equitable, and legally defensible foundation for collaborative medical intelligence The path forward requires continued interdisciplinary collaboration among computer scientists developing security and privacy technologies, healthcare professionals articulating clinical requirements and safety constraints, legal scholars and ethicists navigating regulatory frameworks and ethical principles and policymakers creating governance structures that enable beneficial innovation while protecting vulnerable populations

Declarations

Funding: The authors declare that no funds, grants, or other support were received during the preparation of this manuscript Conflict of interest: The authors have no relevant financial or non-financial interests to disclose Ethics approval: This study utilizes the MIMIC-IV dataset, which is a deidentified public dataset Availability of data and materials: The data that support the findings of this study are available from PhysioNet (MIMIC-IV), but restrictions apply to the availability of these data, which were used under license for the current study Code availability: The source code developed for this study is publicly available in the GitHub repository: https://github.com/rodrigoronner/TBFL-EHR-Framework Authors’ contributions: R. p. 5

Where the paper says it

The work by (Kubach and Roßnagel 2024) addresses the broader challenges of Self-Sovereign Identity adoption, noting that while SSI provides a theoretically sound framework for user-centric identity control and privacy preservation, its practical integration into automated machine learning workflows remains largely unexplored. The authors highlight specific technical and operational challenges, including the complexity of credential issuance processes, the need for standardized verification protocols, and the computational overhead associated with cryptographic identity verification.

T., L.A., and R.A.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Identity-First Security Layer Implementation: The system will enforce trust in identity by requiring cryptographic verification of Verifiable Credentials before allowing any model update submission to the aggregation server, directly neutralizing Sybil attacks through a deterministic gatekeeper mechanism.

  2. Proactive Poisoning Attack Neutralization: By mandating on-chain verification for every model update, the framework ensures that only entities with verified institutional legitimacy and regulatory compliance can contribute, which prevents malicious actors from injecting corrupted gradients designed to degrade global model performance.

  3. Clinical Safety Assurance through Conservative Prediction: The system will maintain a recall metric of 0.890 and an AUC-ROC of 0.954 by using the SMOTETomek algorithm locally to create balanced training sets, preventing the trivial solution of predicting survival for all cases.

  4. Robust Convergence Under Heterogeneity: The integration of the FedProx algorithm with a proximal term (mu = 0.01) ensures that the global model achieves stable convergence and maintains high performance even when aggregating data from substantially different patient population characteristics.

  5. Economic Viability through Content-Addressed Storage: Model updates will be uploaded to IPFS, where only a compact CID (typically 46 bytes) is submitted on-chain, ensuring that the total operational cost remains approximately 18 for 100 training rounds across multiple institutions regardless of model complexity.

  6. Multi-Modal Data Integration Capability: The architecture can be extended to support multi-modal federated learning by allowing separate models to be trained for different data types (e.g., imaging, text) and then fusing them at a late stage while maintaining identity verification for each modality's contribution.

Abstract

Federated learning allows hospitals to train a shared clinical model without exchanging patient records, but its aggregation assumes that each participant is a distinct institution. In an open federation, an adversary can register many fake identities (a Sybil attack) and outweigh honest hospitals. Byzantine-robust aggregation mitigates such attacks only while the adversary is a minority. We ask whether binding participation to accountable institutional identity closes that gap, and at what cost. We present TBFL, a framework in which an accountable authority issues Verifiable Credentials to the Decentralized Identifiers of hospitals, and a smart-contract registry on a peer-to-peer ledger admits model updates only from credentialed senders. We evaluated a prototype on in-hospital mortality prediction with MIMIC-IV records split across simulated hospitals, comparing the identity gate with federated averaging, three robust aggregators, norm clipping, and a similarity-based defense. Robust aggregators held against a Sybil minority, and norm clipping only against random vectors; all failed when half the participants were Sybils trained on real records with flipped labels. The similarity-based defense stopped mutually similar Sybils but yielded to perturbed ones. The gate rejected all unauthorized submissions and maintained attack-free performance for every strategy up to 60% Sybils. It did not stop credentialed hospitals that flipped their own labels, whereas the gate followed by the coordinate-wise median withstood both. On a local node, on-chain checks added about 1% to training time; fees made training impractical on a public mainnet but affordable on a rollup. Identity gating complements robust aggregation rather than replacing it.

Sources

Related papers