Strong Transferable Adversarial Attacks via Ensembled Asymptotically Normal Distribution Learning

Zhengwei Fang, Rui Wang, Tao Huang, Liping Jing

Abstract: Strong adversarial examples are crucial for evaluating and enhancing the robustness of deep neural networks. However, the performance of popular attacks is usually sensitive, for instance, to minor image transformations, stemming from limited information ─ typically only one input example, a handful of white-box source models, and undefined defense strategies. Hence, the crafted adversarial examples are prone to overfit the source model, which hampers their transferability to unknown architectures. In this paper, we propose an approach named Multiple Asymptotically Normal Distribution Attacks (MultiANDA) which explicitly characterize adversarial perturbations from a learned distribution. Specifically, we approximate the posterior distribution over the perturbations by taking advantage of the asymptotic normality property of stochastic gradient ascent (SGA), then employ the deep ensemble strategy as an effective proxy for Bayesian marginalization in this process, aiming to estimate a mixture of Gaussians that facilitates a more thorough exploration of the potential optimization space. The approximated posterior essentially describes the stationary distribution of SGA iterations, which captures the geometric information around the local optimum. Thus, MultiANDA allows drawing an unlimited number of adversarial perturbations for each input and reliably maintains the transferability. Our proposed method outperforms ten state-of-the-art black-box attacks on deep learning models with or without defenses through extensive experiments on seven normally trained and seven defense models.

Introduction

Adversarial examples—inputs crafted with small, imperceptible perturbations to deceive deep neural networks—are vital for assessing model robustness. However, existing attacks often overfit the specific architecture of the source model, leading to poor "transferability" to other models or those with defensive mechanisms. This paper introduces MultiANDA, a novel approach that treats adversarial perturbations as a learned distribution, significantly enhancing their ability to fool diverse and unknown target models.

Methodology

The core of the proposed method lies in two algorithms: ANDA and MultiANDA.

  1. ANDA (Asymptotically Normal Distribution Attack): This algorithm leverages the asymptotic normality property of Stochastic Gradient Ascent (SGA). By using a constant learning rate and data augmentations, the optimization trajectory settles into a stationary Gaussian distribution. ANDA captures the mean and covariance of this process to "map" the local geometry of the loss landscape.
  2. MultiANDA: This extends the approach by employing a Deep Ensemble strategy. By running multiple ANDA processes from different initializations, it estimates a Mixture of Gaussians. This allows the attack to explore multiple basins of the loss landscape simultaneously, capturing a more diverse and robust set of adversarial directions.

Key Insights and Visual Analysis

A critical observation is that MultiANDA perturbations are not just random noise; they are geometrically aware. Visualizations demonstrate that these perturbations tend to cluster around semantically informative regions of an image, such as the head of a bird or the petals of a flower. This focus on features that all high-performing classifiers rely on is a key reason for the method's high transferability across different architectures.

Visual comparison of adversarial perturbations. MultiANDA perturbations (right) are more concentrated on semantic features compared to baselines like VMI-FGSM (left).

Experimental Results

The researchers evaluated MultiANDA against 14 different models (7 normally trained and 7 with advanced defenses) using the ImageNet-1k dataset. MultiANDA consistently outperformed ten state-of-the-art (SOTA) black-box attacks.

As shown in the performance data below, when attacking from a ResNet-50 source, MultiANDA achieved an average success rate of 95.2% against normally trained models, compared to 78.4% for the previous SOTA, VMI-FGSM. The performance gap was even more pronounced against defense models; for instance, against the High-level representation Guided Denoiser (HGD), MultiANDA maintained an 82.3% success rate, while most baseline attacks plummeted to below 40%.

Attack success rates (%) of various methods against normally trained models. MultiANDA consistently achieves the highest transferability across different source and target architectures.

Conclusion

By treating adversarial perturbations as a learned distribution rather than a single point, MultiANDA provides a more robust and transferable framework for evaluating AI security. While the ensemble approach increases computational requirements, its efficiency and effectiveness set a new benchmark for black-box adversarial attacks. Future work will focus on developing defenses that align more closely with human visual perception to mitigate such semantically-targeted vulnerabilities.

Summary updated

Abstract

Strong adversarial examples are crucial for evaluating and enhancing the robustness of deep neural networks. However, the performance of popular attacks is usually sensitive, for instance, to minor image transformations, stemming from limited information ─ typically only one input example, a handful of white-box source models, and undefined defense strategies. Hence, the crafted adversarial examples are prone to overfit the source model, which hampers their transferability to unknown architectures. In this paper, we propose an approach named Multiple Asymptotically Normal Distribution Attacks (MultiANDA) which explicitly characterize adversarial perturbations from a learned distribution. Specifically, we approximate the posterior distribution over the perturbations by taking advantage of the asymptotic normality property of stochastic gradient ascent (SGA), then employ the deep ensemble strategy as an effective proxy for Bayesian marginalization in this process, aiming to estimate a mixture of Gaussians that facilitates a more thorough exploration of the potential optimization space. The approximated posterior essentially describes the stationary distribution of SGA iterations, which captures the geometric information around the local optimum. Thus, MultiANDA allows drawing an unlimited number of adversarial perturbations for each input and reliably maintains the transferability. Our proposed method outperforms ten state-of-the-art black-box attacks on deep learning models with or without defenses through extensive experiments on seven normally trained and seven defense models.
BibTeX
@inproceedings{fang2022strong,
  author = {Zhengwei Fang and Rui Wang and Tao Huang and Liping Jing},
  title = {{Strong Transferable Adversarial Attacks via Ensembled Asymptotically Normal Distribution Learning}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2022},
  doi = {10.48550/arxiv.2209.11964},
}
Figure