Patch-based Adversarial Attack against DNNs

Deep neural networks (DNNs) have revolutionized machine learning with their remarkable capabilities in various applications. However, their vulnerability to adversarial attacks, where subtle perturbations can lead to misclassification, raises significant concerns regarding their reliability and security. In this paper, we study adversarial attacks by implementing a sophisticated patch-based methodology to assess the vulnerability of DNNs. Our approach involves freezing the weights of trained classification models and introducing a patch parameter that is strategically placed on images from diverse datasets such as MNIST, CIFAR-100, and a transportation dataset comprising 43 categories. The primary objective is to evaluate the impact of these patches on DNNs’ classification accuracy, particularly focusing on their robustness against white-box and black-box attacks. Our results demonstrate high attack success rates of up to 56.63% in compromising DNNs’ classification accuracy.