Exploiting Certified Defences to Attack Randomised Smoothing

Andrew C. Cullen, Paul Montague, Shijie Liu, Sarah M. Erfani, Benjamin I. P. Rubinstein

Abstract

In guaranteeing that no adversarial examples exist within a bounded region, certification mechanisms play an important role in demonstrating the robustness of neural networks. In this work we ask: Could certifications have any unintended consequences, through exposing additional information about certified models? We answer this question in the affirmative, demonstrating that certifications not only measure model robustness but also present a new attack surface. We propose Certification Aware Attacks (CAA), that produce smaller adversarial perturbations more than twice as frequently as any prior approach, when launched against certified models. Our attacks achieve an up to 34% reduction in the median perturbation norm, while requiring 90% less computational time than approaches like PGD. That our attacks achieve such significant reductions in perturbation size and computational cost highlights an apparent paradox in deploying certification mechanisms.

The Paradox of Certified Robustness

Adversarial examples—inputs specifically designed to deceive deep learning models—remain a critical security challenge. To address this, researchers developed Certified Robustness. Techniques such as Randomized Smoothing or Interval Bound Propagation (IBP) provide a mathematical guarantee: a "safe bubble" (radius) around an input where no adversarial examples can exist.

While these certificates are intended to increase trust and security, they introduce a significant information asymmetry. If an attacker knows the exact size of this "safe zone," they no longer need to search the interior. Instead, they can use the certificate as a roadmap to find vulnerabilities exactly where the model's defenses are unproven.

Methods: The Certification Aware Attack (CAA)

The authors introduce the Certification Aware Attack (CAA), a two-stage framework that weaponizes safety guarantees:

  1. Step Size Control (The Search): Traditional attacks like Projected Gradient Descent (PGD) take small, uniform steps. CAA uses the model's own certification radius to take "informed jumps." By skipping the certified safe region in a single step, the search lands precisely at the boundary of the unknown, significantly accelerating the discovery of adversarial points.
  2. Refining Adversarial Examples (The Shrink): Once a vulnerability is found, CAA exploits the certification of the adversarial point itself. Since any point within the certified radius of an adversarial example is likely also adversarial, the attack "pulls" the image back toward the original input, minimizing the perturbation norm while staying outside the model's safe zone.
Conceptual overview of the Certification Aware Attack (CAA) showing the search and refinement process.

To handle non-differentiable components like the argmax layer in randomized smoothing, the framework employs a Gumbel Softmax re-parameterization, enabling gradient-based optimization on previously "black-box" parts of the pipeline.

Experimental Evaluation

The framework was evaluated on standard computer vision benchmarks—MNIST, CIFAR-10, and ImageNet—against models defended by Randomized Smoothing and Interval Bound Propagation (IBP).

Key Results

CAA consistently outperformed established attacks like PGD, Carlini-Wagner, and AutoAttack:

  • Efficiency: CAA produced norm-minimizing adversarial examples 74% more often than comparable attacks.
  • Stealth: It reduced the median perturbation norm (the "size" of the attack) by more than 10% on average, and up to 52% on complex ImageNet data.
  • Success Rate: CAA reached success rates of 90-97% while maintaining significantly smaller perturbation sizes than competitors.
Success rate of CAA (red/blue lines) compared to other attacks across different perturbation radii.

Conclusion: The Security Implications of Transparency

The most striking conclusion of this research is that the act of certifying a model paradoxically reduces its security. By publishing a certificate, a defender hands the attacker a map of the model's blind spots. The authors suggest that to mitigate these risks, certifications themselves should be treated as highly confidential data. Future work may explore "noisy" certifications or alternative ways to provide safety guarantees without exposing the underlying robustness landscape to exploitation.

Summary updated

Abstract

In guaranteeing that no adversarial examples exist within a bounded region, certification mechanisms play an important role in demonstrating the robustness of neural networks. In this work we ask: Could certifications have any unintended consequences, through exposing additional information about certified models? We answer this question in the affirmative, demonstrating that certifications not only measure model robustness but also present a new attack surface. We propose \emph{Certification Aware Attacks}, that produce smaller adversarial perturbations more than twice as frequently as any prior approach, when launched against certified models. Our attacks achieve an up to $34\%$ reduction in the median perturbation norm (comparing target and attack instances), while requiring $90 \%$ less computational time than approaches like PGD. That our attacks achieve such significant reductions in perturbation size and computational cost highlights an apparent paradox in deploying certification mechanisms. We end the paper with a discussion of how these risks could potentially be mitigated.
BibTeX
@inproceedings{cullen2023exploiting,
  author = {Andrew C. Cullen and Paul Montague and Shijie Liu and Sarah M. Erfani and Benjamin I. P. Rubinstein},
  title = {{Exploiting Certified Defences to Attack Randomised Smoothing}},
  booktitle = {International Conference on Learning Representations},
  year = {2023},
  url = {https://openreview.net/forum?id=AeTl9sbF-VT},
}
Figure