Fig. 5From: Perturbing BEAMs: EEG adversarial attack to deep learning models for epilepsy diagnosingIn GPBEAM-DE without perturbation overflow, the perturbation added on one electrode are limited below \(\epsilon\) (left). In GPBEAM-DE with Perturbation overflow, the perturbation added on one electrode could be a little bigger at first (middle), but then the perturbation over \(\epsilon\) is distributed equally to all other electrodes (right)Back to article page