Fig. 5
From: On the role of deep learning model complexity in adversarial robustness for medical images
Decision boundary projection of standard trained CBR-LargeT and Resnet50 for Chest X-ray, Dermoscopy & OCT datasets. Unperturbed samples were projected on column 1 (no attack) and perturbed samples were projected on column 2 (PGD attack) for (a, b), respectively. The more complex decision boundary in (b) resulted in samples that were closer to the decision boundary in the projected space which increased medical image DNN sensitivity to input perturbations