Demiguise Attack: Crafting Invisible Semantic Adversarial Perturbations with Perceptual Similarity

Yajie Wang, Shangbo Wu, Wenyi Jiang, Shengang Hao, Yu-an Tan, Quanxin ZhangIJCAI 2021 · DOI

TL;DR

This paper introduces "Demiguise Attack," a novel method for generating adversarial examples that are imperceptible to humans while being highly effective at fooling deep neural networks. By leveraging "Perceptual Similarity" as a metric, Demiguise Attack manipulates semantic information within images, leading to more robust and transferable attacks compared to traditional methods.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are notoriously vulnerable to adversarial examples – subtly modified inputs designed to cause misclassification. Current adversarial attacks often rely on constraining perturbations within tight l_p norms (like l_2 or l_inf). While these norms aim to keep perturbations invisible, they often result in high-frequency noise that, while mathematically small, can still be visually noticeable to humans. This creates a trade-off: stronger attacks often lead to more perceptible perturbations. Furthermore, these l_p-norm-based attacks can be less effective against black-box models or models with defenses like noise reduction filters. The core problem this paper addresses is the creation of adversarial perturbations that are both highly effective at fooling DNNs and truly imperceptible to human vision, moving beyond the limitations of pixel-wise similarity metrics.

Methods

The key innovation is the use of Perceptual Similarity as a metric for crafting adversarial perturbations. Perceptual Similarity, derived from deep features of pre-trained CNNs (specifically VGG in this work), aims to better mimic human judgments of image similarity than traditional l_p norms.

The authors propose Demiguise Attack, a general strategy that integrates Perceptual Similarity into existing adversarial attack frameworks. They demonstrate this with Demiguise-C&W, which combines the Carlini & Wagner (C&W) attack's optimization objective with a perceptual loss function based on Perceptual Similarity. This allows the attack to optimize for both adversarial effectiveness (fooling the classifier) and perceptual imperceptibility simultaneously.

The general optimization procedure involves a joint optimization of an adversarial loss (L_adv) and a perceptual distance loss (L_dist). The paper also shows how Demiguise Attack can be extended to other gradient-based attacks (like MI-FGSM) and decision-based attacks (like HopSkipJumpAttack), enhancing their performance. A key aspect is that Demiguise Attack can leverage not just the distance but also the gradient information from the Perceptual Similarity network, enabling its integration into various attack paradigms.

Datasets

The experiments were conducted using the ImageNet dataset. The authors randomly selected 1000 images from 10 classes for evaluation.

Key Insights and Ideas

  • Perceptual Similarity is Superior to l_p Norms: Traditional l_p norms fail to capture human perception of image similarity, leading to visually detectable noise. Perceptual Similarity, by leveraging deep features, provides a more human-aligned metric for crafting perturbations.
  • Semantic Manipulation for Imperceptibility: Demiguise Attack manipulates semantic information within images rather than just pixel values. This allows for larger, more meaningful perturbations that are still imperceptible because they align with how humans understand image content.
  • Simulating Real-World Variations: The perturbations generated by Demiguise Attack can mimic natural illumination and contrast changes, suggesting that DNNs might have inherent blind spots related to these variations.
  • Additive and Universal Strategy: Demiguise Attack is presented as a general strategy that can be combined with existing attacks, significantly boosting their effectiveness, robustness, and transferability.

Key Results

  • Enhanced Imperceptibility: Demiguise-C&W generated perturbations that were significantly more imperceptible than those from C&W (l_2), C&W-PSNR, and C&W-SSIM, while achieving 100% fooling rates.
  • Improved Robustness: Demiguise Attack showed a 3-5% increase in fooling rates compared to l_p-based C&W when faced with defenses like JPEG compression and binary filters.
  • Increased Transferability: Combining Demiguise Attack with MI-FGSM resulted in a 10-30% increase in transferability, and further incorporating input diversity (DI-MI-FGSM) pushed transfer-based fooling rates to over 90%.
  • Semantic Meaning in Perturbations: When perturbations were passed through classifiers alone, Demiguise-C&W achieved classification ratios of around 80%, significantly higher than l_2-based C&W (around 40%), indicating that Demiguise perturbations carry semantic meaning.
  • Competitive Performance: Demiguise Attack ranked third in the Security AI Challenger VI Track II: Unrestricted Adversarial Attacks on ImageNet at CVPR2021.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, the use of deep feature extraction for Perceptual Similarity likely adds computational overhead compared to simpler l_p norm calculations.
  • Future Work: The authors suggest that the ability of Demiguise Attack to simulate natural phenomena like illumination changes implies the potential for developing semantic defenses for DNNs, which they plan to explore further. They also highlight the promising implications for exposing blind spots in DNNs.

Summary updated

BibTeX
@inproceedings{wang2021demiguise,
  author = {Yajie Wang and Shangbo Wu and Wenyi Jiang and Shengang Hao and Yu-an Tan and Quanxin Zhang},
  title = {{Demiguise Attack: Crafting Invisible Semantic Adversarial Perturbations with Perceptual Similarity}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/430},
}
Figure