Abstract

Recent studies have uncovered the susceptibility of Deep Neural Networks (DNNs) to manipulation through small perturbations applied to input vectors. This project focuses on analyzing a specific attack scenario, where only a single pixel can be modified. To address this, we propose a novel approach utilizing differential evolution (DE) to generate one-pixel adversarial perturbations. Our method operates as a black-box attack, requiring minimal adversarial information, and demonstrates the ability to deceive a wider range of network types due to the inherent characteristics of DE. Our results reveal that, 67.97% of the natural images in the Kaggle CIFAR-10 test dataset can be perturbed to target at least one different class by modifying just one pixel, achieving an average confidence of 74.03%. This attack sheds light on the vulnerability of current DNNs to low-dimensional attacks, presenting a distinctive perspective in the field of adversarial machine learning within an extremely constrained scenario.

Read Project Report for detailed explaination

Project Report.pdf