Table 1 The difference between this study and existing studies of EEG white-box adversarial attacks. In contrast to [13, 23, 24], this paper focuses on the vulnerability of DNN; Compared to [22, 25], in addition to studying the vulnerability of CNNs, this paper also studies the vulnerability of CNN + RNN; Unlike existing studies, this paper generates EEG adversarial samples by perturbing BEAMs, as the input to internal architecture is BEAMs; In addition, this paper examines not only dense attacks, but also sparse attacks
From: Perturbing BEAMs: EEG adversarial attack to deep learning models for epilepsy diagnosing
Related studies | Victim model type | The internal architecture of the victim model | Inputs to internal architecture | Type of attack |
|---|---|---|---|---|
Zhang et al. [13] | Non-DNN | Canonical correlation analysis; Logistic Regression | EEG, EEG frequency | Dense |
Aminifar [23] | SVM | EEG | Dense | |
Meng et al. [24] | Logistic regression | EEG | Dense | |
Zhang and Wu [22] | DNN | CNN | EEG, EEG spectrogram | Dense |
Feng et al. [25] | CNN | EEG | Sparse | |
This paper | CNN; RNN + CNN | BEAMs | Dense; Sparse |