User-wise Perturbations for User Identity Protection in EEG-Based BCIs

TL;DR

This research shows the feasibility of hiding user identity information in EEG data without impacting the primary BCI task information and proposes four types of user-wise privacy-preserving perturbations.

Abstract

Objective: An electroencephalogram (EEG)-based brain-computer interface (BCI) is a direct communication pathway between the human brain and a computer. Most research so far studied more accurate BCIs, but much less attention has been paid to the ethics of BCIs. Aside from task-specific information, EEG signals also contain rich private information, e.g., user identity, emotion, disorders, etc., which should be protected. Approach: We show for the first time that adding user-wise perturbations can make identity information in EEG unlearnable. We propose four types of user-wise privacy-preserving perturbations, i.e., random noise, synthetic noise, error minimization noise, and error maximization noise. After adding the proposed perturbations to EEG training data, the user identity information in the data becomes unlearnable, while the BCI task information remains unaffected. Main results: Experiments on six EEG datasets using three neural network classifiers and various traditional machine learning models demonstrated the robustness and practicability of the proposed perturbations. Significance: Our research shows the feasibility of hiding user identity information in EEG data without impacting the primary BCI task information.

Figures (6)

• Figure 1: Both task classification model and UID model can be trained on the MI EEG dataset. The training of task classification model requires EEG trails and task labels (left hand, right hand, tongue, and both feet), and the training of UID model needs EEG trails and user labels.
• Figure 2: Illustration of user identity protection in EEG-based MI classification. The UID model trained on identity-unlearnable training EEG data cannot distinguish user identities from the original (unperturbed) test data. In contrast, the BCI task classifier trained on these identity-unlearnable EEG data still performs well on the original test data. This indicates that identity-unlearnable EEG data still contains learnable task information, but the user identity information is hidden and cannot be learned by the machine learning model.
• Figure 3: An original EEG trial and its perturbed counterpart with (a) RAND, (b) SN, (c) EMIN, and (d) EMAX.
• Figure 4: Test BCAs of different UID models under AT on (a) MI1; (b) MI2; (c) MI3; (d) MI4; (e) ERN; and (f) P300.
• Figure 5: Test BCAs of the UID models under different data preprocessing/transformations on (a) MI1, (b) MI2, (c) MI3, and (d) MI4. SL: Surface Laplacian; TS: Temporal Shift; TR: Temporal Recombination. SL, which needs at least four channels, was not performed on MI4, as it has only three channels.