Improving Transferability of Adversarial Examples with Virtual Step and Auxiliary Gradients

Ming Zhang, Xiaohui Kuang, Hu Li, Zhendong Wu, Yuanping Nie, Gang ZhaoIJCAI 2022 · DOI

Abstract

Deep Neural Networks (DNNs) are susceptible to adversarial examples, but transferring these examples from a known substitute model to an unknown black-box target remains a significant challenge. This paper presents a novel optimization strategy to boost transferability by addressing the issue of local optima and overfitting. We introduce two key techniques: the Virtual Step, which employs larger, unclipped movements to explore a broader optimization space, and Auxiliary Gradients, which leverage multiple labels to diversify the adversarial path. Our proposed method, VA-I-FGSM, achieves substantial improvements in attack success rates, outperforming existing state-of-the-art transfer-based attacks by 12% to 28% across various target models and defenses.

Introduction

Adversarial attacks are typically categorized into white-box and black-box scenarios. In black-box settings, attackers often rely on the transferability of adversarial examples crafted on a local substitute model. However, standard iterative attacks like the Iterative Fast Gradient Sign Method (I-FGSM) often overfit the substitute model's specific decision boundaries, leading to poor performance on different target architectures. This research aims to break these "black boxes" by encouraging the optimization process to find more generalizable adversarial regions.

Methodology

The core of our approach lies in two synergistic techniques integrated into the iterative gradient-based attack framework.

Virtual Step and Auxiliary Gradients

Traditional attacks take small steps and apply clipping at each iteration to ensure the perturbation remains within a predefined $\epsilon$-ball. This frequent clipping can trap the optimization in poor local maxima. The Virtual Step technique allows the optimizer to take larger steps without intermediate clipping. Clipping is only applied at the final iteration, enabling the search process to traverse a wider landscape and potentially reach more effective adversarial regions.

Standard attacks utilize only the gradient of the ground-truth label to move the image "away" from the correct class. We propose the use of Auxiliary Gradients derived from multiple random labels. By simultaneously moving away from the true label and navigating relative to auxiliary labels, the attack path becomes more diverse. This diversity increases the likelihood that the resulting adversarial example falls into a classification region that is consistently misclassified across different model architectures. This aligns with the "Distance Hypothesis": adversarial examples are more likely to transfer if they are pushed into feature space regions far removed from the original ground-truth class.

Conceptual illustration of the VA-I-FGSM attack path using virtual steps and auxiliary gradients.

Experimental Results

We evaluated the proposed VA-I-FGSM method using a 1000-image subset from ImageNet against a variety of models, including Inception-v3, Xception, Inception-ResNet-v2, and ResNet-152-v2, as well as several defense-oriented models.

Performance Gains

The integration of virtual steps and auxiliary gradients yielded significant improvements. For instance, when transferring an attack from Inception-v3 to Xception, the success rate increased from a baseline of 13.8% to 41.6%. Across all tested configurations, VA-I-FGSM consistently outperformed state-of-the-art baselines like DI²-FGSM and TI²-FGSM by margins of 12% to 28%. The method also proved highly effective against models using advanced defenses like JPEG compression and denoising.

Visual Fidelity

Despite the more aggressive optimization path enabled by the virtual step, the final adversarial images remain visually indistinguishable from their benign counterparts. The perturbations are strictly constrained within the standard threshold of $\epsilon = 16/255$, ensuring the attacks remain stealthy to human observers.

Visual comparison of adversarial examples generated by different methods, demonstrating high visual fidelity.

Conclusion

By combining virtual steps and auxiliary gradients, we have demonstrated a powerful method for enhancing the transferability of adversarial examples. This approach effectively mitigates overfitting to substitute models and provides a more robust tool for evaluating the security of black-box DNN systems. Future work will explore the application of these techniques to targeted attacks and further optimize the computational efficiency of auxiliary gradient calculations.

Summary updated

BibTeX
@inproceedings{zhang2022improving,
  author = {Ming Zhang and Xiaohui Kuang and Hu Li and Zhendong Wu and Yuanping Nie and Gang Zhao},
  title = {{Improving Transferability of Adversarial Examples with Virtual Step and Auxiliary Gradients}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/227},
}
Figure