Task-Preserving EEG Anonymization Using Latent Feature Masking

The growing use of electroencephalogram (EEG) data in deep learning offers the potential for revolutionary new applications while also raising significant privacy concerns. Given the heavily idiosyncratic nature of EEG data, a bad actor might easily exploit public datasets to identify specific subjects and, hence, infer potentially sensitive information about their mental state, health status, etc. In the paper, we propose, a method to selectively remove identity features from EEG data, while preserving the utility of the dataset for its intended purpose. We test SAFE against three distinct threat models using four popular EEG classification datasets. Results demonstrate that SAFE provides strong protection against adversarial actors without seriously compromising task-relevant features. Our approach addresses the important yet largely unexplored need for robust protections and safety against privacy threats to EEG data while supporting data sharing, enabling scientific and clinical advances.