New CNN model boosts EEG-based brain-computer interface security against adversarial attacks

A new study from researchers Md Fahimul Kabir Chowdhury and Gahangir Hossain tackles the overlooked security vulnerabilities in EEG-based brain-computer interfaces (BCIs). While most prior work focused on boosting classification accuracy, the team warns that BCIs are susceptible to adversarial attacks—tiny, deliberate disturbances to EEG signals that can cause misdiagnosis. To counter this, they designed a lightweight custom Convolutional Neural Network (CNN) architecture that balances robustness with computational efficiency. The model was benchmarked against three novel EEG-specific CNNs—EEGNet, DeepConvNet, and SleepEEGNet—on two distinct EEG datasets under gradient-based adversarial attack scenarios (e.g., FGSM, PGD).

Experimental results show the proposed model consistently maintained higher classification accuracy under adversarial perturbations compared to all baselines, demonstrating significantly improved robustness. The lightweight nature of the architecture makes it practical for real-time BCI applications where computational resources are limited. Accepted and presented at IEEE World AI IoT Congress 2026, this research highlights the urgent need to embed security into BCI design from the ground up. By proving that simpler models can still defend against sophisticated attacks, the work opens the door to safer neural interfaces for medical diagnostics, assistive technologies, and beyond.