Motor-imagery brain-computer interfaces (BCIs) use EEG signals that encode a person's intended movement. But those signals also carry a less obvious payload: a stable, subject-specific signature. According to a new preprint on arXiv, this identifying structure persists across recording sessions, meaning that separately released EEG datasets from the same person could be linked by an adversary.
The preprint (arXiv:2610.15626) proposes a Riemannian suppression approach to counter this cross-session identity leakage. Rather than relying on conventional anonymisation, the method operates on the geometric properties of EEG covariance matrices, which are a standard representation in Riemannian BCI pipelines.
If the approach holds up, it would address a growing privacy concern as EEG datasets are shared for research and commercial development. The key risk is not just re-identification of a single recording, but the ability to link multiple recordings from the same individual across time. The Riemannian framing is a distinctive response because it targets the structure that makes the leakage persistent, rather than treating each session in isolation.