acceptodds
Withdrawn from ICLR 2027

Riemannian Suppression of Cross-Session Identity Leakage in Motor-Imagery EEG

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Electroencephalography (EEG) recorded for motor-imagery (MI) brain-computer interfaces (BCIs) carries subject-identifying structure that persists across sessions, allowing separately released sessions from the same person to be linked. Existing EEG identity protections use learned or additive perturbations and operate on representations other than the covariance geometry underlying MI pipelines. We ask whether cross-session identity leakage can be suppressed directly in this covariance representation without learning a privacy model. We introduce a Riemannian transform that equalizes two subject-conditioned geometric statistics: the Riemannian mean covariance and the tangent-space scatter around it, with two optional band-specific refinements. Ablations show that mean alignment and scatter whitening carry nearly all of the suppression, while the optional stages contribute small, dataset-dependent refinements. Across two public datasets and five attackers spanning four covariance-space representations, the worst attacker falls from 90.8% to 15.6% on BCI Competition IV-2a and from 90.8% to 9.9% on OpenBMI, at a cost of only 7.0 and 2.1 points of MI decoding accuracy relative to the 8–30 Hz input, and no aggregation rule exceeds chance by more than two points at any batch size. The transform requires no learned privacy model, surrogate network, or gradient-based training, and is fitted on one session and applied unchanged to a held-out later session.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, brain-computer interface, motor imagery, privacy, re-identification, Riemannian geometry, spatial covariance, cross-session linkage, training-free

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Dz3xOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates cross-session identity leakage in motor-imagery EEG systems and proposes a Riemannian covariance-space transformation to suppress subject identity while preserving task-relevant information. Although the results are promising and the paper is well written and easy to follow, the paper has very limited machine learning novelty. The proposed method is mainly a reuse of well-developed Riemannian alignment under distribution shifts to extract domain-invariant information for cross-subject decoding. In my opinion, these are two questions within the same spirit, so the paper's contribution to the ML community is insufficient.

Critical strengths

S1: The identity leakage issue exists in real-world applications, and adapting Riemannian alignment shows promising results.

Critical weaknesses

W1: This paper has very limited novelty related to the machine learning community. Its main contribution is to introduce the well-developed Riemannian alignment for distribution shifts in a new setting. As Riemannian alignment under distribution shifts aims to extract domain-invariant information, the identity issue is straightforward to solve. While the optional stages contribute little, the main contribution is again from Riemannian alignment.
I don't see any potential impact of this paper on the machine learning community at ICLR.

W2: This paper studied a general identity leakage problem but only tested on motor imagery, ignoring other BCI tasks that share the same identity leakage issue (e.g., SSVEP) [1].

[1] Meng, Lubin, et al. "User identity protection in EEG-based brain–computer interfaces." IEEE transactions on neural systems and rehabilitation engineering 31 (2023): 3576-3586.

Questions for authors

What's the contribution of the paper other than introducing Riemannian alignment to the identity leakage setting?

Why focus only on motor imagery data? Is there any specific reason not to consider other BCI tasks that could share the same issue?

Reviewer_DbvcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a training-free geometric transform that aligns subject covariance means and tangent-space scatter to reduce cross-session EEG identification. Across two datasets, it substantially suppresses five evaluated covariance-based attackers while retaining useful motor-imagery decoding performance. Other tested attacks retain substantial identity information.

Critical strengths
  1. Simple method with informative ablations. The transform requires no learned privacy model, and the ablations distinguish the contributions of mean alignment and scatter adjustment.
  2. Substantial empirical evaluation. Session-1-only parameter selection, multiple attackers, alignment baselines, and explicit disclosure of residual leakage make the measured privacy–utility trade-off informative.
Critical weaknesses
  1. The covariance-space claim needs clarification. Band-power identification remains at 72.7%/33.6%. If these features are channel powers over the same band and window, they derive from covariance diagonals and cannot simply be classified as outside covariance space. The evidence supports suppression of selected attackers, not protection of the representation generally.
  2. Near-chance session linkage is insufficiently established. Appendix H does not identify the score-producing classifier or demonstrate aggregation of the strongest defended attackers. Moreover, fitting on the reference session leaves linkage between two independently held-out sessions untested.
Questions for authors
  1. How exactly are band-power features computed, and what distinguishes them from covariance-diagonal features?
  2. Which classifier supplies Appendix H’s scores? Does aggregation of the strongest defended attackers also remain near chance?
  3. Can you evaluate linkage with both reference and query data excluded from transform fitting, or explicitly restrict the conclusions to calibration-to-future linkage?
Reviewer_XfL6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies identity leakage across sessions in motor imagery EEG and its tradeoff with decoding utility.

The authors propose a geometric transformation that aligns covariance means and tangent space scatter, then reconstructs time domain signals. They fit the transformation on one session and apply it to another without training a privacy model.

Experiments on BCI-IV-2a and OpenBMI show that identification by the strongest evaluated covariance attacker falls from 90.8% to 15.6% and 9.9%, respectively. However, band power and EEGNet retain substantial identification accuracy in the released signals, limiting the scope of protection.

Critical strengths
  1. The problem is well motivated, and the finding that Euclidean Alignment still leaves 69.0% worst case identification (Table 3) is useful in its own right.

  2. The proposed transform is simple and training free, and the operating point is chosen on Session 1 alone, with blind selection costing only 0.4 points (App. J).

  3. The experiments are thorough and candid, including stage ablations (Table 2), subject level tests and an explicit report of where the transform stops (Sec. 6.4).

Critical weaknesses
  1. My main concern is whether the evidence supports the claim of covariance space protection. Band power identification remains at 72.7% on BCI-IV-2a in Sec. 6.4, despite its close relationship to the covariance diagonal. If diagonal identification succeeds, covariance attackers’ performance near chance may reflect classifier failure rather than identity removal. The defended tangent space SVM on BCI-IV-2a gives exactly 11.1% in Table 1, Table 11 (γ_scw≥0.35) and every Table 13 row, consistent with constant predictions.

  2. The headline cost is measured after filtering, which understates the total loss on BCI-IV-2a. Table 1 shows losses relative to raw input of 8.7 points for FBCNet and 12.9 for FBCSP. With only nine subjects, nonsignificant utility losses for EEGNet and DeepConvNet (p=0.16 and 0.20, App. K) do not demonstrate preserved utility, whereas App. K calls the asymmetry between significant reductions in attacker accuracy and nonsignificant utility losses “the substantive point”.

  3. The paper does not show whether its geometry improves privacy at matched utility. Sec. 7 attributes suppression to increased dispersion, but Table 3 lacks a comparable noise control and omits the RPA rotation step. The paper acknowledges that RMA and SCW resemble Riemannian recentering and second order alignment, respectively (lines 107 and 200–201). The contribution therefore centers on privacy application and evaluation, making matched utility controls essential.

  4. The robustness evaluation needs broader coverage of alternative features and aggregation. App. G refits the defense for each subband, rather than attacking the fixed full band release with subband features, while App. H leaves the classifier used for the reported aggregation results unspecified.

  5. I am not convinced that the verification result demonstrates removal of identity information. Table 6 reports FRR=0.989: nearly every pair is rejected at this threshold, so EER or ROC-AUC would better assess residual discrimination.

Questions for authors
  1. Could the authors compare attacks using diagonal and full covariance on the same release, with scaling and tuning restricted to training data? How is band power estimated and normalized? If this comparison shows that identity information is removed from the covariance representation, I would raise my score.

  2. What is the privacy benefit of SCW over RMA plus noise at matched utility relative to raw input, using the same selection procedure across utility tolerances?

  3. Have the authors tested reversed distance rankings, with the matching rule chosen on the training session? Table 9 reports 0.4±0.7% at N=25 on OpenBMI, but 1.9% with all trials.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.