Title: Enhancing Security in Brain-Computer Interfaces

Abstract:

The rapid evolution of electroencephalogram (EEG)-based brain-computer interfaces (BCIs) has been largely driven by advancements in machine learning. While previous studies have predominantly prioritized improving classification precision, there has been a notable lack of attention regarding the security and resilience of these systems. Recent evidence suggests that EEG-based BCIs are vulnerable to adversarial attacks; specifically, subtle, carefully designed disturbances can lead to misdiagnosis. Consequently, assessing how models withstand such perturbations is essential for guaranteeing their safe and effective deployment.

This paper introduces a lightweight, custom Convolutional Neural Network (CNN) architecture designed to examine the adversarial robustness of EEG-based BCIs. We evaluated this proposed method using two distinct EEG datasets, comparing its performance against three specialized CNN models tailored for EEG data: EEGNet, DeepConvNet, and SleepEEGNet. These comparisons were conducted within the context of gradient-based adversarial attack scenarios.

Experimental results demonstrate that our proposed model maintains superior classification performance under adversarial perturbations compared to the baseline models, signaling enhanced robustness. These outcomes underscore the promise of lightweight architectural designs in bolstering the reliability of EEG-based BCI systems when facing adversarial threats.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC