Towards Interpreting and Utilizing Symmetry Property in Adversarial Examples

Shibin Mei, Chenglong Zhao, Bingbing Ni, Shengchao YuanAAAI 2023 · OpenReview

Abstract

Deep Neural Networks (DNNs) are notoriously vulnerable to adversarial examples—natural images modified by tiny, invisible perturbations. In this work, we identify a "symmetry property" in adversarial scenarios by viewing attacks at a fine-grained level. We propose a new metric, Attack Proportion, to measure the percentage of adversarial examples misclassified between specific classes. We observe that these proportions are highly unbalanced, with classes showing vulnerability to specific "symmetry pairs." By explicitly hardening the decision boundaries between these pairs during training, we develop a defense mechanism that outperforms current state-of-the-art robust models.

Introduction

While most research on adversarial robustness focuses on the overall Attack Success Rate, this paper investigates the fine-grained relationships between classes. Current defenses often treat all class errors as equal. However, in reality, certain classes are much more likely to be confused with one another (e.g., "cat" and "dog" vs. "cat" and "ship"). By ignoring this uneven distribution of vulnerability in the feature space, standard training fails to create a sufficiently large "support margin" between visually similar classes, leaving them open to adversarial manipulation.

The Symmetry Property

We introduce Attack Proportion ($P(i \to j)$) to quantify the likelihood of class $i$ being misclassified as class $j$ under attack. Our analysis reveals the Symmetry Property: the observation that $P(i \to j) \approx P(j \to i)$. When this mutual confusion is at its peak, the classes are dubbed a Maximum Symmetry Pair.

Figure 1: Visualization of the symmetry property. Arrows highlight high attack proportions between specific class pairs, suggesting that adversarial vulnerability is deeply tied to the underlying geometry of the feature space.

As shown in the empirical data, adversarial confusion is not random. Maximum symmetry pairs often consist of visually similar objects. The model's failure to learn truly discriminative features for these pairs is what adversarial attacks exploit.

Figure 3: Empirical distribution of attack proportions for MNIST (Digit 3) and CIFAR-10 (Airplane). The data demonstrates that for any given class, adversarial misclassifications are concentrated in a few specific target classes.

Symmetry-Aware Defense Strategy

To address these vulnerabilities, we propose a Symmetry-Aware Loss Function. Instead of standard softmax cross-entropy, this approach treats class weights as "prototypes" in the feature space and applies two core constraints:

  1. Inter-class Constraint: This explicitly penalizes similarity between the weights of different classes. Crucially, it uses an adjustable parameter ($\zeta$) to force an even larger angular margin between identified "maximum symmetry pairs."
  2. Intra-class Constraint: This uses Mean Square Error (MSE) to ensure features of the same class stay close to their designated prototype.

By identifying which class boundaries are the "weakest links" (the symmetry pairs), we can apply stronger regularization specifically where it is needed most.

Experimental Evaluation

The proposed defense was evaluated on standard benchmarks including CIFAR-10, CIFAR-100, SVHN, and TinyImageNet. The method consistently outperformed established methods like TRADES and MART across multiple attack types (FGSM, PGD, AutoAttack).

  • CIFAR-10 (ResNet34): Under a PGD50 attack, the method achieved 49.62% accuracy, surpassing MART (48.19%) and TRADES (48.18%).
  • SVHN: The method reached 59.78% accuracy under PGD50, a significant improvement over MART (58.13%).
  • Black-Box Robustness: The defense also proved highly effective against "black-box" attacks, outperforming prior methods on metrics like NES and SimBA.

Conclusion and Future Work

The symmetry property provides a new lens through which to understand and mitigate adversarial vulnerability. While the current study uses static pairs identified before training, future work may explore dynamic updates to symmetry pairs during the training process. Additionally, scaling these fine-grained regularization techniques to massive datasets like ImageNet remains an important challenge for future research.

Summary updated

Abstract

In this paper, we identify symmetry property in adversarial scenario by viewing adversarial attack in a fine-grained manner. A newly designed metric called attack proportion, is thus proposed to count the proportion of the adversarial examples misclassified between classes. We observe that the distribution of attack proportion is unbalanced as each class shows vulnerability to particular classes. Further, some class pairs correlate strongly and have the same degree of attack proportion for each other. We call this intriguing phenomenon symmetry property. We empirically prove this phenomenon is widespread and then analyze the reason behind the existence of symmetry property. This explanation, to some extent, could be utilized to understand robust models, which also inspires us to strengthen adversarial defenses.
BibTeX
@inproceedings{mei2023towards,
  author = {Shibin Mei and Chenglong Zhao and Bingbing Ni and Shengchao Yuan},
  title = {{Towards Interpreting and Utilizing Symmetry Property in Adversarial Examples}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  url = {https://openreview.net/forum?id=GAT5ZwnuGh},
}
Figure