A Systematic Review of Security Privacy and Provenance in EEG-Based Brain-Computer Interfaces with a Blockchain-Based Reference Architecture

Technical Summary: A Systematic Review of Security, Privacy, and Provenance in EEG-Based Brain-Computer Interfaces with a Blockchain-Based Reference Architecture

Problem Statement
Neural data generated by Brain-Computer Interfaces (BCIs), particularly those utilizing Electroencephalography (EEG) for motor imagery, presents a unique security challenge: unlike passwords or physical biometric templates, neural patterns cannot be revoked if compromised. As BCIs transition from controlled laboratories to clinical rehabilitation and consumer applications, the lack of systematic analysis regarding the security, privacy, and provenance of these data pipelines poses significant risks to patient safety and regulatory compliance.

The review identifies a critical gap: while deep learning architectures (CNNs, transformers, federated models) have revolutionized motor imagery classification, they introduce substantial security surfaces. These include vulnerability to adversarial perturbations that could mislead classifiers controlling prosthetics, membership inference attacks that expose participant identity, and a reliance on public benchmark datasets (e.g., BCI Competition IV, PhysioNet) that lack cryptographic integrity guarantees, leaving them susceptible to undetectable data poisoning. Despite the theoretical suitability of blockchain for providing immutable, auditable provenance, the intersection of blockchain technology and EEG-based BCI systems has received negligible attention.

Methodology
This study is a PRISMA 2020-compliant systematic review. The authors registered a protocol on the Open Science Framework (OSF) and conducted a multi-database search across five sources: IEEE Xplore, PubMed, Scopus, SciSpace, and Google Scholar/arXiv.

• Search Scope: The search covered literature from January 2017 to February 2026, focusing on motor imagery tasks in healthy adults or clinical populations (stroke, spinal cord injury, ALS).
• Eligibility Criteria (PICOS): The review included studies using deep learning methods (CNNs, RNNs, transformers, GNNs) and/or blockchain/privacy mechanisms applied to EEG-BCI pipelines. It excluded invasive recordings (ECoG), non-motor paradigms (P300, SSVEP), and traditional machine learning without security analysis.
• Selection Process: From an initial pool of 729 records, 678 unique records were screened after deduplication. Following title/abstract screening and full-text review, 64 high-quality papers were selected for in-depth analysis.
• Quality Assessment: A modified JBI Critical Appraisal Tool was used to score studies on ten criteria, including dataset justification, validation methods, and overfitting prevention.

Key Contributions

1. Quantification of the Provenance Gap: The review provides the first multi-database quantification of the blockchain research gap in BCI. Only 4 out of 729 reviewed papers (<1%) addressed blockchain-based provenance for EEG-BCI systems. Notably, the search of IEEE Xplore (67 records) returned zero results for blockchain implementations in motor imagery BCI.
2. Security Vulnerability Characterization: The authors systematically analyzed four classes of security vulnerabilities introduced by dominant deep learning architectures:
   • Adversarial Vulnerability: Susceptibility of CNNs and transformers to imperceptible input perturbations that alter predictions.
   • Privacy Risks: Vulnerability to Membership Inference Attacks (MIAs) and model inversion, where adversaries can determine training set membership or reconstruct sensitive attributes from model weights.
   • Data Poisoning: The lack of cryptographic integrity in public datasets allows for supply-chain attacks where modified data silently corrupts downstream models.
   • Federated Learning Risks: While federated learning reduces raw data exposure, it remains vulnerable to gradient inversion attacks without formal differential privacy guarantees.
3. NeuroChain Reference Architecture: Based on the identified gaps and vulnerabilities, the authors propose a conceptual "NeuroChain" reference architecture. This architecture is not a specific software implementation but a set of design requirements including:
   • Acquisition-stage Integrity: Hash anchoring of raw EEG data before preprocessing to prevent poisoning.
   • Provenance Generation: Signed records at every pipeline transition.
   • Privacy-Preserving Federation: Integration of formal differential privacy at the aggregation step to counter gradient inversion.
   • Patient-Controlled Consent: Revocable smart contracts allowing patients to audit who has trained models on their data.
   • Latency Constraints: Off-chain storage and inference to meet the <300 ms latency requirement for real-time BCI control, with only hashes anchored on-chain.

Results

• Publication Trends: The 64 analyzed papers were predominantly published between 2021 and 2025. The most common venues were IEEE TNSRE and the Journal of Neural Engineering.
• Architectural Performance:
  • Within-Subject: State-of-the-art models like RNN-LSTM achieved up to 98% accuracy on BCI IV-2a, while TCFormer reached 96.27% on HGD.
  • Cross-Subject/Session: Performance drops significantly in cross-subject scenarios (e.g., 51–89%), with EEGSym showing strong cross-subject results (88.6±9.0%).
  • Methodological Quality: Common issues included insufficient sample size justification (65%), lack of independent test sets (38%), and missing hyperparameter reporting (42%).
• Blockchain Implementation Status: The four identified blockchain studies were found to be conceptually limited. None addressed motor imagery specifically, none reported latency impact on BCI performance, and none integrated with mainstream BCI frameworks (e.g., MNE-Python, OpenViBE).

Significance and Claims
The paper asserts that the current landscape of EEG-based BCI research is characterized by a "stark quantitative imbalance" between the proliferation of deep learning architectures and the near-total absence of provenance and security mechanisms. The authors claim that the identified gap is not an artifact of database selection but a genuine field-wide deficiency confirmed across five independent indexing sources.

The significance of this work lies in its derivation of the NeuroChain architecture, which translates abstract security needs into concrete design requirements for future systems. The authors emphasize that achieving clinical translation for BCI systems—particularly for neurorehabilitation in stroke and ALS patients—requires addressing these security and privacy gaps. They argue that without immutable audit logs, cryptographic dataset integrity, and patient-controlled consent, the deployment of BCIs in safety-critical environments remains premature. The review concludes that future research must prioritize real-time latency evaluation of blockchain integration, certified adversarial robustness, and the development of cryptographically secured dataset registries.