Securing ResNet50 against Adversarial Attacks: Evasion and Defense using BIM Algorithm

Deep neural networks, such as ResNet50, have shown remarkable performance in image classification tasks. However, susceptibility to adversarial attacks, where small perturbations to input images can result in misclassifications, is a concern. The BIM algorithm is a popular technique for generating adversarial examples. The objective of this research is to explore the vulnerability of ResNet50 DNNs trained on the ImageNet Stubs dataset to evasion attacks using the BIM algorithm. Additionally, the effectiveness of various defense strategies, including adversarial training, defensive distillation, and input transformations, is examined. The results reveal that ResNet50 DNNs are vulnerable to BIM-based evasion attacks and that these defenses can enhance their robustness. Overall, this work underscores the importance of defending against BIM-based evasion attacks to ensure the security and reliability of ResNet50 DNNs.