Abstract
State-of-the-art machine learning models frequently misclassify inputs that have been perturbed in an adversarial manner. Adversarial perturbations generated for a given input and a specific classifier often "transfer" to other models. This paper provides a theoretical and geometric explanation for this phenomenon, proving that for linear classifiers and certain two-layer ReLU networks, there exist universal adversarial directions that can simultaneously fool multiple models trained on similar data.
Introduction
Adversarial examples—small, often imperceptible perturbations that cause machine learning models to misclassify data—are a well-known security vulnerability. A particularly puzzling aspect of these attacks is transferability: an adversarial perturbation designed to fool Model A often successfully fools Model B, even if Model B has a different architecture or was trained on different data. While transferability has been observed empirically for years, enabling "black-box" attacks, the underlying theoretical reasons were not fully understood. This paper bridges that gap by providing a geometric framework to explain why these shared vulnerabilities exist.
The Geometry of Adversarial Directions
The authors analyze the geometry of decision boundaries to identify "adversarial directions"—vectors that, when added to an input, push it across the decision boundary for many classifiers at once.
Linear Analysis
For any set of linear classifiers that correctly separate a dataset, the max-margin SVM solution ($w^*$) provides a "universal" direction. Specifically, moving in the direction of $-yw^*$ (where $y$ is the label) is guaranteed to eventually misclassify the point for any valid linear separator of that data. This suggests that transferability is a structural property of the models' decision boundaries rather than a fluke of specific training instances.
ReLU Network Analysis
The analysis extends beyond linear models to two-layer ReLU networks. By assuming non-negative weights on the final layer, the authors prove that the decision regions for these networks remain convex. This convexity allows for the existence of transferable adversarial directions similar to the linear case. If a classifier's "wrong" label region is convex, there is a wide "corridor" of directions that lead into it, making it easy for different models to share that vulnerability.
Empirical Validation and Complexity Asymmetry
The theoretical findings were validated using the MNIST dataset. The authors trained 50 neural networks (NNs) and 50 Support Vector Machines (SVMs) on different random subsets of the training data.
A key insight is Complexity Asymmetry: adversarial directions for complex models (ReLU networks) are more likely to transfer to simpler models (linear classifiers) than vice versa. This suggests that as models become more complex, their adversarial vulnerabilities "encompass" those of simpler models.
In empirical tests, NN-generated attacks were highly effective against SVMs, dropping accuracy significantly. Conversely, SVM-generated attacks were significantly less effective against NNs, with accuracy remaining above 75% even at higher perturbation levels, as shown in the figure above.
Conclusion
This work demonstrates that adversarial transferability is not merely an empirical curiosity but a predictable consequence of the geometry of decision boundaries. By shifting the focus from individual adversarial points to universal adversarial directions, the authors provide a robust framework for understanding cross-model vulnerabilities. Future work may explore these properties in deeper architectures where non-convexity presents greater theoretical challenges.
Summary updated