Investigating Membership Inference Attacks Against CNN Models for BCI Systems

Area of Science:

• Neuroscience
• Computer Science
• Data Privacy

Background:

• Deep learning (DL) adoption in healthcare raises privacy concerns.
• Brain-computer interfaces (BCIs) increasingly use DL, specifically Convolutional Neural Network (CNN) classifiers for Electroencephalogram (EEG) data.

Purpose of the Study:

• To investigate privacy vulnerabilities of CNN classifiers for EEG data in BCIs.
• To analyze Membership Inference Attacks (MIA) in the context of heterogeneous datasets and spatial-temporal design choices.

Main Methods:

• Empirical analysis of MIA susceptibility to training dataset specifics (participant number, demographics).
• Investigation of MIA susceptibility to CNN specifics (architecture, regularization).
• Comparison of MIA effectiveness on EEG data versus image and tabular datasets.

Main Results:

• An adversary with limited knowledge can compromise participant privacy via MIA on EEG data.
• Training on diverse datasets improves overall privacy but risks memorization for underrepresented groups.
• Regularization is less effective against MIA for EEG data CNNs; model depth/width has no impact.

Conclusions:

• DL-based BCIs are susceptible to privacy breaches through MIA.
• Privacy-aware BCI system development is crucial, considering dataset diversity and model characteristics.
• Further research is needed to mitigate MIA risks in DL-based BCI applications.