A Frequency-Based Analysis Method to Improve Adversarial Robustness of Neural Networks for EEG-Based Brain-Computer Interfaces

With the recent advancement of the Brain-Computer Interface (BCI), deep neural networks (DNNs) for Electroencephalogram (EEG) classifications have gained considerable research attention. However, these DNN networks are vulnerable to adversarial attacks, thus leading to the misclassification of EEG signals by decreasing the model accuracy. Herein, an easy-to-use and automatic method is proposed to determine whether unseen EEG samples are normal or adversarial, and improve DNN robustness from the frequency perspective. On two publicly available EEG datasets, the mean detection rates of 88.33% and 83.25% respectively is achieved against the adversarial attack. The results validate the efficiency of the hereby-proposed method in being generalized to different attack methods and perturbation levels with no need for retraining.