Adaptive Pixel Resilience: A Novel Defence Mechanism Against One-Pixel Adversarial Attacks on Deep Neural Networks

This paper presents a groundbreaking analysis of the One Pixel Attack, an insidious adversarial threat that challenges the robustness of state-of-the-art deep neural networks (DNNs). We delve into the intricate mechanics of this deceptively simple yet potent attack, which can cause misclassification by altering just a single pixel in an image. Our research not only unravels the technical underpinnings of the One Pixel Attack but also introduces Adaptive Pixel Resilience (APR), a novel defence mechanism that significantly enhances DNN robustness against this threat. Through extensive experimentation on the CIFAR10 and ImageNet datasets, we demonstrate the remarkable efficacy of APR. Our method substantially outperforms existing defence strategies, setting a new benchmark in adversarial robustness while maintaining competitive clean accuracy. The paper offers several key contributions