Abstract
Strong adversarial examples are the keys to evaluating and enhancing the robustness of deep neural networks. The popular adversarial attack algorithms maximize the non-concave loss function using the gradient ascent. However, the performance of each attack is usually sensitive to, for instance, minor image transformations due to insufficient information (only one input example, few white-box source models and unknown defense strategies). Hence, the crafted adversarial examples are prone to overfit the source model, which limits their transferability to unidentified architectures. In this paper, we propose Multiple Asymptotically Normal Distribution Attacks (MultiANDA), a novel method that explicitly characterizes 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 apply the ensemble strategy on this procedure to estimate a Gaussian mixture model for a better exploration of the potential optimization space. Drawing perturbations from the learned distribution allow us to generate any number of adversarial examples for each input. The approximated posterior essentially describes the stationary distribution of SGA iterations, which captures the geometric information around the local optimum. Thus, the samples drawn from the distribution reliably maintain the transferability. Our proposed method outperforms nine state-of-the-art black-box attacks on deep learning models with or without defenses through extensive experiments on seven normally trained and seven defence models.
Introduction
Adversarial attacks—subtle, human-imperceptible perturbations that trick AI models into making wrong predictions—are the "stress tests" of the deep learning world. However, most attacks suffer from a major flaw: they are too specific. An attack that fools a ResNet-50 model often fails against an Inception model. This lack of transferability suggests that current attack methods overfit the quirks of a single "source" model rather than finding the fundamental vulnerabilities of neural networks.
The core challenge in adversarial machine learning is black-box transferability. In a real-world scenario, an attacker doesn't know the architecture or weights of the target model. To circumvent this, they craft an attack on a "substitute" model. If the attack is transferable, it will work on the unknown target too. Current state-of-the-art (SOTA) attacks often find a single point in the high-dimensional image space that maximizes error, but that point might be a "narrow peak" of vulnerability unique to that specific model. To find "broad valleys" of vulnerability that exist across many models, we need a more robust way to explore the perturbation space.
Methodology
MultiANDA shifts the paradigm from finding a single adversarial image to learning a whole distribution of them. The method consists of two nested algorithms:
- ANDA (Asymptotically Normal Distribution Attack): Instead of standard deterministic gradient ascent, the authors use Stochastic Gradient Ascent (SGA) combined with random data augmentations. They leverage a property of SGA called asymptotic normality: as optimization nears a local optimum, the path it takes describes a Gaussian (Normal) distribution. By tracking the mean and covariance of the gradients, ANDA "maps" the local geometry of the vulnerability.
- MultiANDA: A single Gaussian distribution is unimodal and only looks at one "basin" of vulnerability. MultiANDA uses a Deep Ensemble strategy, running several ANDA processes from different random starting points. This creates a Gaussian Mixture Model (GMM), allowing the attack to explore multiple regions of the optimization space simultaneously.
By sampling from this learned mixture distribution, the attacker can generate an unlimited number of diverse adversarial examples for a single input image, all of which are likely to be effective.
Analysis of Semantic Focus
Visualization of the perturbations generated by MultiANDA reveals that they naturally concentrate on the "semantic" areas of an image (e.g., a bird's head) rather than random pixels. This explains why they are so effective at confusing different models' decision-making processes, as they target the core features used for classification across architectures.
Experimental Results
The researchers evaluated MultiANDA using the ImageNet-1k dataset against 7 normally trained models (ResNet, Inception, VGG) and 7 advanced defense models. They even tested transferability to Vision Transformers (ViTs) and CLIP.
MultiANDA consistently outperformed ten SOTA black-box attacks (including MI-FGSM, DIM, and VMI-FGSM). On normally trained models, MultiANDA achieved success rates often exceeding 90% in black-box settings. The most impressive results were against defense models; for example, when attacking the "HGD" defense model using a ResNet-50 source, MultiANDA achieved an 82.3% success rate, while the next best SOTA method (VMI-FGSM) only reached 37.7%.
Conclusion
MultiANDA demonstrates that a distribution of perturbations is inherently more robust than a single point. By capturing the geometric information around local optima, the attack becomes less sensitive to minor model differences. While the primary limitation is the computational cost of running multiple optimization loops, the significant gains in transferability—especially against defended models—mark a substantial advancement in adversarial research. Future work may explore using these learned distributions to improve model robustness through enhanced adversarial training.
Summary updated