Adversarial and Clean Data Are Not Twins

Abstract

Adversarial attack has cast a shadow on the massive success of deep neural networks. Despite being almost visually identical to the clean data, the adversarial images can fool deep neural networks into wrong predictions with very high confidence. In this paper, however, we show that we can build a simple binary classifier separating the adversarial apart from the clean data with accuracy over 99%. We also empirically show that the binary classifier is robust to a second-round adversarial attack. In other words, it is difficult to disguise adversarial samples to bypass the binary classifier. Further more, we empirically investigate the generalization limitation which lingers on all current defensive methods, including the binary classifier approach. And we hypothesize that this is the result of intrinsic property of adversarial crafting algorithms.

Cited by in corpus (9)

• Adversarial Examples: Opportunities and Challenges
• Adversarial Example Detection for DNN Models: A Review and Experimental Comparison
• Adversarial Machine Learning in Image Classification: A Survey Towards the Defender's Perspective
• Detection of Face Recognition Adversarial Attacks
• PatchCensor: Patch Robustness Certification for Transformers via Exhaustive Testing
• Two Souls in an Adversarial Image: Towards Universal Adversarial Example Detection using Multi-view Inconsistency
• AdvFAS: A robust face anti-spoofing framework against adversarial examples
• Perturbation Analysis of Gradient-based Adversarial Attacks
• HoneyModels: Machine Learning Honeypots