On the Robustness of the CVPR 2018 White-Box Adversarial Example Defenses

Anish Athalye, Nicholas CarliniarXiv 2018 · OpenReview

Abstract

Neural networks are known to be vulnerable to adversarial examples. In this note, we evaluate the two white-box defenses that appeared at CVPR 2018 and find they are ineffective: when applying existing techniques, we can reduce the accuracy of the defended models to 0%.

Introduction

Adversarial examples—inputs with subtle, intentional perturbations designed to deceive machine learning models—pose a significant security risk. At the CVPR 2018 conference, two prominent defense mechanisms were proposed to mitigate this vulnerability:

  1. Pixel Deflection: This method randomly replaces selected pixels with nearby neighbors and applies a denoising operation to disrupt adversarial patterns.
  2. High-Level Representation Guided Denoiser (HGD): This approach utilizes a trained neural network to denoise inputs before they reach the classifier.

While these defenses claimed robustness against "oblivious" attackers (those unaware of the defense), this study evaluates their effectiveness under a realistic, rigorous white-box threat model, where the attacker has full knowledge of the defense strategy and adapts accordingly.

Methodology

The evaluation utilizes a strict white-box threat model on the ImageNet validation set. Rather than developing new attacks, the authors demonstrate that standard, existing techniques are sufficient to bypass these defenses. To ensure a rigorous test, adversarial perturbations were constrained to a strict $L_\infty$ bound of $\epsilon = 4/255$, which is more restrictive than the bounds used in the original defense papers.

Attacking Pixel Deflection

Pixel Deflection relies on non-differentiable operations, which typically break standard gradient-based attacks by causing "gradient masking." To overcome this, the authors employed Backward Pass Differentiable Approximation (BPDA). BPDA approximates gradients through non-differentiable layers, allowing for effective optimization of adversarial perturbations even through randomized preprocessing.

Attacking HGD

Since the HGD denoiser is a standard differentiable neural network prepended to the classifier, it can be treated as part of an end-to-end classification pipeline. The authors applied Projected Gradient Descent (PGD) directly through the combined denoiser and classifier without modification.

Results and Discussion

The evaluation reveals that both defenses are completely vulnerable to adaptive white-box attacks. The authors successfully reduced the accuracy of models protected by both Pixel Deflection and HGD to exactly 0%.

Adversarial examples generated against Pixel Deflection and HGD defenses. The top row shows original images, while the middle and bottom rows show successful targeted attacks.

Key Findings

  • High Targeted Success Rates: For Pixel Deflection, the BPDA attack achieved a 97% success rate in generating targeted adversarial examples (defined as hitting the target label 9 out of 10 times due to the defense's randomness). For HGD, the end-to-end PGD attack achieved a perfect 100% targeted success rate.
  • The Illusion of Non-Differentiability: Pixel Deflection's reliance on non-differentiable preprocessing provides a false sense of security. Adaptive techniques like BPDA easily circumvent the resulting gradient masking.
  • Vulnerability of Differentiable Preprocessors: Differentiable defenses like HGD offer no protection against end-to-end gradient-based attacks; they simply become another layer for the optimizer to traverse.

Conclusion

The results underscore the inadequacy of evaluating adversarial defenses against oblivious threat models. Security claims based on "gradient masking" or simple differentiable preprocessing often fail when faced with an adaptive attacker. This work highlights that researchers proposing new defenses must prioritize rigorous white-box evaluation using strong, adaptive attacks to ensure meaningful progress in adversarial robustness.

Summary updated

Abstract

Neural networks are known to be vulnerable to adversarial examples. In this note, we evaluate the two white-box defenses that appeared at CVPR 2018 and find they are ineffective: when applying existing techniques, we can reduce the accuracy of the defended models to 0%.
BibTeX
@misc{athalye2018robustness,
  author = {Anish Athalye and Nicholas Carlini},
  title = {{On the Robustness of the CVPR 2018 White-Box Adversarial Example Defenses}},
  year = {2018},
  eprint = {1804.03286},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1804.03286},
}
Figure