Generating Adversarial Attacks On Medical Images Using Classical And Transform Domain Techniques

Abstract

The most powerful and common machine learning models used for manydifferent image analysis tasks like 3D analysis, image retrieval, image classification, object detection are known as deep neural networks (DNNs). Theyhave achieved performance level near to human level. Based on the successof DNNs on natural images (e.g. captured images from natural scenes likeimagenet and CIFAR-10), they have become very popular for tasks such asmedical image processing, organ/landmark localization, diagnosis of cancer, diabetic retinopathy detection and Covid-19 identification. Despite the surgein the development of smart intelligent systems by inculcating deep struc-tured networks, the vulnerabilities of such smart systems may impede by amere involvement of an adversarial perturbation. The adversarial perturba-tions may not be detected by the naked eye, but they can be seen by deepneural network classifiers. In this study, a novel methodology is proposedthat crafts perturbations by using a transform domain techniques and classi-cal approach like FGSM on Covid19 images thereby fooling the state-of-artDNN models