Architecture-Agnostic Iterative Black-Box Certified Defense Against Adversarial Patches

Di Yang, Yihao Huang, Qing Guo, Felix Juefei-Xu, Ming Hu, Yang Liu, Geguang Pu

TL;DR

Researchers have developed IBCD (Iterative Black-box Certified Defense), the first architecture-agnostic defense that provides provable security against adversarial patches without knowing the patch's size or location. By iteratively estimating patch dimensions through a novel double-masking search, IBCD bridges the gap between theoretical "white-box" certified defenses and unpredictable real-world attack scenarios.

What Problem Does This Paper Address?

Adversarial patches are a nightmare for computer vision. Unlike subtle "noise" attacks, these are visible, localized regions (like a sticker on a stop sign) that can completely flip a model's prediction.

To combat this, researchers use certified defenses, which provide a mathematical guarantee that a model will remain robust within certain bounds. However, current state-of-the-art (SOTA) certified defenses have a major "white-box" flaw: they assume the defender already knows the size of the adversarial patch. In a real-world deployment—say, an autonomous car encountering a vandalized sign—the system has no way of knowing if a patch is 5x5 pixels or 50x50. If the size estimate is wrong, the certification fails. This paper tackles the "patch size unknown" problem to make certified defenses actually practical.

Methods

The authors propose a two-stage framework called IBCD:

  1. Patch Size Estimation: This is the core innovation. Since they don't know the patch size, they treat it as a search problem.
  2. Double Masking: They use a "one-mask" then "double-mask" approach. If a mask successfully covers the adversarial patch, the model's prediction becomes consistent. By observing when the model's output stabilizes vs. when it fluctuates, they can "box in" the likely size of the patch.
  3. Sliding Space Optimization: Searching every possible mask position is computationally expensive. IBCD uses a "coarse-to-fine" strategy. Once a large mask is found to cover the patch, the algorithm only searches for smaller masks within that specific region, drastically reducing the search space.
  4. Certified Defense: Once the size is estimated, they plug that value into existing SOTA white-box defenses (like PatchCleanser or Derandomized Smoothing) to calculate the final robust prediction.

Crucially, the method is architecture-agnostic. It treats the underlying neural network as a black box, making it compatible with both traditional Convolutional Neural Networks (CNNs) and modern Vision Transformers (ViTs).

Datasets

The authors evaluated IBCD on two benchmark datasets:

  • CIFAR-10: 10,000 test images (32x32 resolution).
  • ImageNet: 100,000 test images (224x224 resolution).

They tested these against two representative architectures: ResNet50 (CNN) and ViT (Transformer).

Key Insights and Ideas

  • The Consistency Signal: The "aha moment" here is that you don't need to see the patch to measure it. You only need to see how the model's uncertainty changes as you mask different parts of the image. Consistency is a proxy for "the threat has been neutralized."
  • The 1/4 Rule: The algorithm operates under the assumption that an adversarial patch won't exceed 25% of the image area. This is a pragmatic constraint; if a patch covers half the image, the original object is likely unrecognizable anyway.
  • Provable Bounds: The paper provides formal proofs (Theorems 1 and 2) showing that their iterative search will "certifiably" find an interval containing the true patch size, ensuring the defense remains mathematically sound.

Key Results

  • Accuracy Trade-offs: While there is a drop in accuracy compared to white-box methods (where the size is perfectly known), the "fluctuation rate" remains acceptable. For example, on ImageNet-ViT, the clean accuracy drop was less than 4.5%.
  • Efficiency Gains: The Sliding Space Optimization is highly effective. In tests on CIFAR-10, it reduced the number of required searches by roughly 50% compared to a vanilla search, making the defense significantly faster.
  • Cross-Architecture Success: The method performed consistently across both ResNet and ViT, proving that the defense doesn't rely on specific gradient information or model layers.

Limitations and Future Work

  • Baseline Performance: The authors honestly note that the "upper bound" of their defense is limited by the performance of the underlying white-box methods. Currently, certified patch defenses still struggle with low absolute certified accuracy on complex datasets like ImageNet.
  • Two-Stage Latency: Because the method requires an iterative search before the final defense, it is slower than a standard forward pass.
  • Future Direction: The authors suggest moving toward a one-stage end-to-end method that could estimate size and defend simultaneously, potentially using a specialized neural network to speed up the estimation phase.

Summary updated

Abstract

The adversarial patch attack aims to fool image classifiers within a bounded, contiguous region of arbitrary changes. To address this problem in a trustworthy way, the certified patch defense methods are proposed. However, the state-of-the-art certified defenses inevitably needed to access the size of the adversarial patch, which is unreasonable and impractical in real-world attack scenarios. To improve the feasibility of the architecture-agnostic certified defense in a black-box setting, we propose a novel two-stage Iterative Black-box Certified Defense method, termed IBCD. In the first stage, it estimates the patch size in a search-based manner by evaluating the size relationship between the patch and mask with pixel masking. In the second stage, the accuracy results are calculated by the existing white-box certified defense methods with the estimated patch size. The experiments conducted on two popular model architectures and two datasets verify the effectiveness and efficiency of IBCD.
BibTeX
@inproceedings{yang2024architecture,
  author = {Di Yang and Yihao Huang and Qing Guo and Felix Juefei-Xu and Ming Hu and Yang Liu and Geguang Pu},
  title = {{Architecture-Agnostic Iterative Black-Box Certified Defense Against Adversarial Patches}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  doi = {10.1109/icassp48485.2024.10448145},
}
Figure