User-wise perturbations for user identity protection in EEG-based BCIs.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Electroencephalogram (EEG)-based brain-computer interfaces (BCIs) offer direct brain-computer communication.
• While BCI accuracy is widely studied, the ethical implications, particularly data privacy, are less explored.
• EEG signals contain sensitive private information, including user identity, emotions, and health status, necessitating robust protection.

Purpose of the Study:

• To investigate methods for protecting sensitive user identity information within EEG signals used for BCIs.
• To develop and evaluate privacy-preserving techniques that prevent the unlearning of identity from EEG data.
• To ensure that privacy enhancements do not negatively impact the core functionality of BCIs.

Main Methods:

• Proposed four novel user-wise privacy-preserving perturbation techniques: random noise, synthetic noise, error minimization noise, and error maximization noise.
• Applied these perturbations to EEG training data to obscure user-specific identity markers.
• Evaluated the effectiveness of perturbations using multiple neural network classifiers and traditional machine learning models across six diverse EEG datasets.

Main Results:

• Demonstrated that the proposed user-wise perturbations successfully render user identity information in EEG data unlearnable.
• Confirmed that these perturbations do not interfere with or degrade the performance of the primary BCI task.
• Validated the robustness and practical applicability of the developed privacy-preserving methods through extensive experiments.

Conclusions:

• It is feasible to effectively hide user identity information within EEG data.
• Privacy-preserving perturbations can be implemented without compromising essential BCI task performance.
• This research lays the groundwork for more ethical and secure development of BCI technologies.