BCIwiki (bciwiki.com) — According to arXiv, Sheikh Samit Muhaimin and Aarthy Nagarajan uploaded a preprint on October 11, 2026, examining EEG privacy in motor imagery brain-computer interfaces (BCIs). The preprint has not been peer reviewed. It addresses cross-session identity leakage: electroencephalography (EEG) retains person-identifying structure across recording sessions, potentially allowing separately released recordings from the same person to be linked.
The researchers introduce a Riemannian transform that operates directly on covariance representations used in motor imagery pipelines, which describe how signals vary together. It equalizes two subject-conditioned geometric statistics: the Riemannian mean covariance and the tangent-space scatter around that mean. Two optional band-specific refinements are also included. The transform requires no learned privacy model, surrogate network, or gradient-based training. It is fitted on one session and applied unchanged to a held-out later session. Ablation tests show that mean alignment and scatter whitening account for nearly all of the suppression, while the optional steps add small, dataset-dependent refinements.
Tests covered two public datasets and five attackers spanning four covariance-space representations. The strongest attacker's accuracy fell from 90.8% to 15.6% on BCI Competition IV-2a and from 90.8% to 9.9% on OpenBMI. Compared with the 8-30 Hz input, motor imagery (MI) decoding accuracy decreased by 7.0 and 2.1 percentage points, respectively. At every tested batch size, no aggregation rule exceeded chance-level identification accuracy by more than two percentage points. These findings describe a trade-off between identity suppression and decoding performance on public datasets, not the elimination of all identity leakage risks.