Differentially Private EEG Feature Anonymization: A Privacy-Utility Case Study in Clinical Neurophysiology
cs.CR · cs.LG · eess.SP
Quick verdict
Case study of Gaussian/Laplace DP on 363-dim EEG features (n=17): candid about Laplace miscalibration, but utility collapses and tiny cohort caps conclusions.
Summary
This case study applies subject-level differential privacy to patient-level EEG feature representations for clinical data sharing. From 122 clinical EEG recordings (17 patients, ~33 channels), the authors extract 363-dimensional patient-level vectors (11 per-channel statistical and band-power features), robustly scale them, apply L2 clipping (C=2.0), and release perturbed versions via an analytic Gaussian mechanism (δ=1e-5, ε∈{0.5–20}, Δ2=4) and coordinate-wise Laplace noise, alongside three described multi-hospital deployment scenarios with HMAC-based linkage. The paper is candid that the implemented Laplace scale (4/ε) is about 19× smaller than the formally required full-vector L1 calibration (76.21/ε), so Laplace results carry no formal DP guarantee, and that robust-scaling statistics fitted on the private data break end-to-end guarantees. Utility collapses: Gaussian correlation peaks at 0.074 and SNR at −20.8 dB even at ε=20, while the LOSO MLP degrades from 0.706 accuracy / 0.545 F1 on original features to 0.294 (Gaussian, all-positive predictions) and 0.235 (Laplace) at ε=10. Main limitations: one noise realization per ε, only 17 patients with 12/5 imbalance, thresholds tuned on the same cohort, no empirical privacy attacks, and no actual federated experiments despite Scenario 3.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Subject-level Gaussian and Laplace perturbation can be integrated into clinical EEG feature-processing workflows, but the mechanism, ε, and sensitivity calibration strongly determine utility.
- Under analytic Gaussian calibration (Δ2=4.0, δ=1e-5), even ε=20 leaves correlation at 0.0743 and SNR at −20.8 dB, i.e., severe utility loss across the entire tested range.
- The implemented Laplace sweep (b=4/ε) preserves more statistical utility than the Gaussian mechanism at equal ε, but is explicitly not a formal full-vector DP release; formal calibration would require b=76.21/ε.
- Downstream epilepsy classification (LOSO MLP) degrades to near-useless at ε=10 (accuracy 0.294 Gaussian, 0.235 Laplace, vs 0.706 on original features).
- Small, imbalanced clinical EEG cohorts (17 patients, 12/5) make DP privacy–utility evaluation particularly difficult.
Strengths
Explicit calibration honesty
The paper separates the implemented Laplace scale b=4/ε from the formally required full-vector scale b=76.21/ε (Eqs. 13–15, Table 1) and states three separate times (Secs. 3.3, 3.3.1, 4.2) that Laplace results carry no formal ε-DP guarantee; it also acknowledges that data-fitted robust scaling breaks end-to-end guarantees.
Clean formal setup
Bounded replace-one subject-level adjacency is explicitly defined (Sec. 3.3), the analytic Gaussian accountant (Balle & Wang [30]) replaces the classical σ formula, and the Related Work carefully distinguishes one-time feature release from DP-SGD accounting.
Honest negative-result reporting
Linkage separated from anonymization
Unusually complete reproducibility table
Weaknesses
Single noise realization per configuration
Guarantee-validity gap from data-fitted scaling
Mechanism comparison confounded by noise magnitude
n=17, 12/5 imbalance, cohort-tuned thresholds
Scheme–evaluation gap
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.11777) is a case study applying subject-level differential privacy (DP) to patient-level EEG feature vectors for multi-hospital data sharing. Three deployment scenarios are proposed — client-side anonymization (Fig. 1), centralized server-side anonymization (Fig. 2), decentralized local training (Fig. 3) — with HMAC-SHA256 linkage (Eq. 1) for exact duplicate detection and a Bloom-filter component for approximate linkage. Central claims: (i) DP perturbation can be integrated into clinical EEG workflows, but utility depends critically on mechanism, ε and sensitivity calibration; (ii) the analytic Gaussian mechanism at Δ2=4.0 destroys statistical utility even at ε=20; (iii) an implemented Laplace sweep (b=4/ε) preserves more utility than Gaussian at equal ε, while formal full-vector Laplace calibration would require b=76.21/ε (Eqs. 13–15). The authors explicitly scope the classification task as a "utility check" and disclaim empirical attack-based privacy testing.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper do?
It applies subject-level differential privacy — Gaussian and Laplace noise calibrated to a 363-dimensional patient-level EEG feature vector — to anonymize clinical EEG data, and measures how much utility (statistical fidelity and epilepsy-classification performance) survives at privacy budgets ε from 0.5 to 20, in a multi-hospital data-sharing framing.
Does the Laplace mechanism in this paper actually satisfy differential privacy?
No. The implemented Laplace scale (4.0/ε) is about 19× smaller than the formally required full-vector L1 calibration (≈76.21/ε). The authors state this openly and derive the formal scales separately, so the Laplace utility numbers carry no formal DP guarantee; only the Gaussian arm is formally calibrated, and even that assumes fixed preprocessing statistics.
How many patients were in the study?
17 patients contributing 122 EEG recordings (~8 GB, ~33 channels). The epilepsy-detection evaluation used 12 non-epilepsy and 5 epilepsy patients with leave-one-subject-out validation on patient-level feature vectors.
Is the code or dataset available?
No public code release is announced. The EEG data are anonymized clinical recordings from the VIKING project (ethics reference 689562) and are not stated to be publicly available, so exact reproduction is not possible from the paper alone.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–11 are raster images; only their captions were extracted, though all key quantitative results are reproduced in Tables 3–7 of the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.