Adversarial Attacks and Defenses of Image Classification: A Comprehensive Survey

Deep neural networks (DNNs) have demonstrated remarkable performance, but significant security vulnerabilities have emerged, first observed in image classification tasks, with significant impact. The core issue of adversarial attacks lies in adversarial examples (AEs), which are crafted by introducing designed perturbations that are imperceptible to the human eye. When these perturbed examples are mixed with clean data, they can cause DNNs to make incorrect classifications. This vulnerability has been observed in various other data-driven domains, exacerbating the insecurity of DNN-based systems. Therefore, a comprehensive analysis of adversarial attacks and defenses in image classification serves as a cornerstone for addressing security issues in other scenarios involving DNNs. This article is motivated by the need to systematically examine the current state of adversarial attacks and defenses in image classification. We summarize the most representative methods in chronological order and re-organizing them into detailed categories. For each method, we analyze its strengths and weaknesses, providing insights into their practical implications. Furthermore, we offer forward-looking perspectives on potential research directions. Unlike existing surveys, we aim to provide valuable references and stimulate the creative solutions for improving the robustness of DNNs in image classification and related application scenarios.