Adversarial Training: A Survey

TL;DR

This survey describes the implementation procedures and practical applications of AT, followed by a comprehensive review of AT techniques from three perspectives: data enhancement, network design, and training configurations.

Abstract

Adversarial training (AT) refers to integrating adversarial examples -- inputs altered with imperceptible perturbations that can significantly impact model predictions -- into the training process. Recent studies have demonstrated the effectiveness of AT in improving the robustness of deep neural networks against diverse adversarial attacks. However, a comprehensive overview of these developments is still missing. This survey addresses this gap by reviewing a broad range of recent and representative studies. Specifically, we first describe the implementation procedures and practical applications of AT, followed by a comprehensive review of AT techniques from three perspectives: data enhancement, network design, and training configurations. Lastly, we discuss common challenges in AT and propose several promising directions for future research.

Figures (2)

• Figure 1: The adversarial perturbation $\epsilon \cdot \text{sign}(\nabla_{x} \ell(x, y;\theta))$ causes the object 'panda' to be misclassified as 'gibbon'.
• Figure 2: A comprehensive overview of adversarial training.