New SW-ProxyCE framework enables adversarial attacks on private EEG models

Researchers have developed a new framework called SW-ProxyCE to address security vulnerabilities in electroencephalography (EEG) foundation models. This framework enables adversarial transfer attacks, where attackers can exploit publicly released EEG encoders to compromise private downstream models without direct access. SW-ProxyCE uses shrinkage-whitened class prototypes to recover decision geometry from limited labeled data, allowing for the generation of transferable adversarial examples. Evaluations across multiple EEG tasks and encoders demonstrated the effectiveness of SW-ProxyCE in transferring adversarial attacks to inaccessible downstream models, outperforming existing methods and highlighting a lack of adversarial robustness in EEG foundation models. AI

IMPACT Highlights potential security risks in the deployment of foundation models for sensitive data like EEG, necessitating further research into robust defenses.

RANK_REASON Academic paper detailing a new method for adversarial attacks on EEG models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• electroencephalography
• GitHub
• Shrinkage-Whitened Proxy Cross-Entropy
• SW-ProxyCE

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →