Hardware-Assisted Privacy-Preserving Multi-Channel EEG Computational Headwear

EEG signals contain highly sensitive information about an individual's mental state, cognitive processes, and health conditions, making privacy preservation crucial. With the rise of commercial headwear capable of capturing EEG signals, developing robust mechanisms for ensuring privacy of such data is imperative. This work aims to protect EEG data privacy in cloud-based processing systems by sending intermediate output after neural network layer splitting to the cloud. We propose a novel holistic Combined Privacy Metric (CPM) that quantifies privacy leakage between raw EEG signals and intermediate outputs. Our study focuses on EEG-based seizure detection using a 1D CNN architecture, achieving accuracy of 96.25%. We evaluate various splitting configurations to optimize the trade-off between privacy preservation and computational efficiency. We find that splitting after the second convolutional layer achieves a CPM of 0.82 with a modest client-side model size of 509kB. This approach significantly enhances EEG data privacy while enabling effective cloud-based analysis, potentially facilitating wider adoption of secure EEG technologies in healthcare and research applications.