Pretrained, Frozen, Still Leaking: Auditing Cross-Encoder Attribute Transfer in EEG Foundation Models
arXiv:2606.09189v1 Announce Type: cross Abstract: EEG foundation-model releases are usually audited one endpoint at a time: raw-reconstruction, membership inference, identity linkage, or DP-SGD on the downstream head. We audit the same released embeddings under all four endpoints jointly, on BIOT, LaBraM, and EEGPT, and show that each single-endpoint audit clears releases that still leak spectral attributes. The decisive evidence is a cross-encoder transfer audit: a single ridge attribute decoder learned from one frozen encoder transfers, via a fitted linear bridge, to held-out-subject test sp
The proliferation of powerful EEG foundation models and the increasing focus on AI safety and privacy are driving the need for more rigorous auditing methods.
This research reveals critical vulnerabilities in current AI auditing practices for brain-computer interfaces, highlighting that even 'frozen' models can leak sensitive biometric data, which has significant implications for user privacy and security.
The standard single-endpoint auditing of EEG foundation models is shown to be insufficient, necessitating a more comprehensive, multi-endpoint and cross-encoder approach to ensure data privacy.
- · AI safety researchers
- · Privacy advocates
- · Developers of robust auditing tools
- · Ethical AI frameworks
- · Developers releasing unaudited EEG foundation models
- · Entities relying on current single-endpoint audits
- · Users of insecure brain-computer interfaces
- · Companies with lax data privacy standards
This discovery will likely lead to an overhaul of auditing standards and practices for foundation models handling sensitive biometric data like EEG.
Increased regulatory scrutiny and potential legal frameworks may emerge to mandate more stringent privacy-preserving techniques in AI development, particularly for health and brain data.
Public distrust in AI systems could grow if these vulnerabilities are not adequately addressed, potentially hindering the adoption of beneficial brain-computer interface technologies.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI