On Global-view Based Defense via Adversarial Attack and Defense Risk Guaranteed Bounds

Trung Le, Anh Tuan Bui, Le Minh Tri Tue, He Zhao, Paul Montague, Quan Hung Tran, Dinh Q. PhungAISTATS 2022 · OpenReview

Abstract

Deep neural networks (DNNs) are susceptible to adversarial attacks, which presents the most severe fragility of the deep learning system. Despite achieving impressive performance, most of the current state-of-the-art classifiers remain highly vulnerable to carefully crafted imperceptible, adversarial perturbations. Recent research attempts to understand neural network attack and defense have become increasingly urgent and important. While rapid progress has been made on this front, there is still an important theoretical gap in achieving guaranteed bounds on attack/defense models, leaving uncertainty in the quality and certified guarantees of these models. To this end, we systematically address this problem in this paper. More specifically, we formulate attack and defense in a generic setting where there exists a family of adversaries (i.e., attackers) for attacking a family of classifiers (i.e., defenders). We develop a novel class of $f$-divergences suitable for measuring divergence among multiple distributions. This equips us to study the interactions between attackers and defenders in a countervailing game where we formulate a joint risk on attack and defense schemes. This is followed by our key results on guaranteed upper and lower bounds on this risk that can provide a better understanding of the behaviors of those parties from the attack and defense perspectives, thereby having important implications to both attack and defense sides. Finally, benefited from our theory, we propose an empirical approach that bases on a global view to defend against adversarial attacks. The experimental results conducted on benchmark datasets show that the global view for attack/defense if exploited appropriately can help to improve adversarial robustness.

Introduction: The Fragility of Deep Learning

Deep Neural Networks (DNNs) are notoriously brittle; small, imperceptible changes to an image—adversarial perturbations—can cause a model to confidently misclassify an object. While many empirical defenses exist, they often lack rigorous theoretical guarantees. This paper identifies a critical "theoretical gap": the lack of understanding regarding the guaranteed limits of attacker performance and defender protection. Most existing methods focus on individual data points. We argue that a global view—understanding how an attack shifts entire class distributions and how these distributions overlap—is essential for building resilient models.

Methodology: A Game-Theoretic Approach

We model the interaction between attackers and defenders as a min-max game, where the attacker maximizes loss and the defender minimizes it.

  1. Multi-distributional $f$-divergence: We developed a mathematical tool to measure divergence between multiple probability distributions simultaneously, quantifying how effectively an attacker "mixes" different classes.
  2. Guaranteed Bounds: We derived mathematical lower bounds (identifying efficient attack strategies) and upper bounds (defining safety limits for defenders).
  3. Global-View (GV) Algorithms: Based on their theory, the authors modified two popular adversarial training methods to incorporate these insights:
  4. GV-PGD: A global-view version of Projected Gradient Descent.
  5. GV-TRADES: A global-view version of the TRADES algorithm.
The global-view approach aims to maintain class separation in both input and latent spaces during adversarial training.

In these "GV" variants, the training process includes a discriminator (similar to those used in GANs). This discriminator helps the model perceive the global distribution, forcing adversarial examples to remain within their original class clusters and away from competing classes in the latent space.

Experimental Setup

We evaluated our methods on standard computer vision benchmarks:

  • MNIST: Handwritten digits (0-9).
  • CIFAR-10: 60,000 color images across 10 classes.
  • Toy2D Dataset: Synthetic clusters used to visualize class intermingling under attack.

Key Insights

  • Intermingling as an Attack Goal: Effective attacks don't just cross decision boundaries; they attempt to make the distribution of adversarial examples from one class indistinguishable from the clean distribution of another.
  • The Defender’s Dual Task: Robustness requires keeping adversarial examples close to their clean counterparts while ensuring global class separation in the model's internal "latent" space.
  • Latent Space Advantage: Performing global-view checks in intermediate layers is more effective than focusing solely on raw input pixels.

Results and Discussion

The "Global-View" variants consistently outperformed standard counterparts across various attacks, including PGD, Auto-Attack (AA), and Boundary & Box (B&B).

Comparison of natural and robust accuracy for standard and Global-View (GV) training methods on MNIST and CIFAR-10.

On CIFAR-10, GV-TRADES achieved a robust accuracy of 53.3% against PGD attacks, surpassing the 51.9% of standard TRADES. Notably, GV methods often improved "natural" accuracy on clean data while simultaneously boosting robustness. For instance, on MNIST, GV-PGD reached 99.5% natural accuracy (vs. 99.4% for PGD-AT) and significantly improved robustness against Auto-Attack (90.5% vs 88.9%). These experimental results align with our derived theoretical bounds, validating the global-view roadmap.

Limitations and Future Work

While effective, the use of GAN-like discriminators increases computational overhead. Future research should explore the scalability of these bounds to larger architectures like Transformers and massive datasets like ImageNet. Additionally, investigating the "global view" against unforeseen attack categories, such as physical-world or patch attacks, remains a priority.

Summary updated

Abstract

It is well-known that deep neural networks (DNNs) are susceptible to adversarial attacks, which presents the most severe fragility of the deep learning system. Despite achieving impressive performance, most of the current state-of-the-art classifiers remain highly vulnerable to carefully crafted imperceptible, adversarial perturbations. Recent research attempts to understand neural network attack and defense have become increasingly urgent and important. While rapid progress has been made on this front, there is still an important theoretical gap in achieving guaranteed bounds on attack/defense models, leaving uncertainty in the quality and certified guarantees of these models. To this end, we systematically address this problem in this paper. More specifically, we formulate attack and defense in a generic setting where there exists a family of adversaries (i.e., attackers) for attacking a family of classifiers (i.e., defenders). We develop a novel class of f-divergences suitable for measuring divergence among multiple distributions. This equips us to study the interactions between attackers and defenders in a countervailing game where we formulate a joint risk on attack and defense schemes. This is followed by our key results on guaranteed upper and lower bounds on this risk that can provide a better understanding of the behaviors of those parties from the attack and defense perspectives, thereby having important implications to both attack and defense sides. Finally, benefited from our theory, we propose an empirical approach that bases on a global view to defend against adversarial attacks. The experimental results conducted on benchmark datasets show that the global view for attack/defense if exploited appropriately can help to improve adversarial robustness.
BibTeX
@inproceedings{le2022global,
  author = {Trung Le and Anh Tuan Bui and Le Minh Tri Tue and He Zhao and Paul Montague and Quan Hung Tran and Dinh Q. Phung},
  title = {{On Global-view Based Defense via Adversarial Attack and Defense Risk Guaranteed Bounds}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2022},
  url = {https://openreview.net/forum?id=lTO_xt7h_O},
}
Figure