The Dark Side of ImageNet: Examining the Impact of Evasion Attacks on Deep Learning

A number of deep learning models have demonstrated great improvement in computer vision, particularly Convolutional Neural Networks (CNNs). However, the robustness and security of these models are being challenged by evasion attacks, where an adversary manipulates the input to cause the model to misclassify it. In this study, projected gradient descent algorithm is used with the line search method to evaluate the impact of evasion attacks on ImageNet. To generate adversarial examples, the PGD with line search algorithm is described first. Following that, ImageNet models are evaluated when attacked by PGD with a line search. This study concludes by discussing current state-of-the-art defenses against PGD with line search attacks and providing directions for future research. The goal of this research is to raise awareness of the security implications of deep learning models trained on ImageNet and to provide insights into how to make these models more robust and secure in the face of PGD with line search attacks.