The Space of Transferable Adversarial Examples

Abstract

Adversarial examples are maliciously perturbed inputs designed to mislead machine learning (ML) models at test-time. They often transfer: the same adversarial example fools more than one model. In this work, we propose novel methods for estimating the previously unknown dimensionality of the space of adversarial inputs. We find that adversarial examples span a contiguous subspace of large (~25) dimensionality. Adversarial subspaces with higher dimensionality are more likely to intersect. We find that for two different models, a significant fraction of their subspaces is shared, thus enabling transferability. In the first quantitative analysis of the similarity of different models' decision boundaries, we show that these boundaries are actually close in arbitrary directions, whether adversarial or benign. We conclude by formally studying the limits of transferability. We derive (1) sufficient conditions on the data distribution that imply transferability for simple model classes and (2) examples of scenarios in which transfer does not occur. These findings indicate that it may be possible to design defenses against transfer-based attacks, even for models that are vulnerable to direct attacks.

15 pages, 7 figures

References in corpus (2)

• Stealing Machine Learning Models via Prediction APIs
• Generating Adversarial Malware Examples for Black-Box Attacks Based on GAN

Cited by in corpus (9)

• Transferable Clean-Label Poisoning Attacks on Deep Neural Nets
• Adversarial Examples, Uncertainty, and Transfer Testing Robustness in Gaussian Process Hybrid Deep Networks
• Extending Defensive Distillation
• Exploring the Space of Black-box Attacks on Deep Neural Networks
• Fooling a Real Car with Adversarial Traffic Signs
• Characterizing Bias in Classifiers using Generative Models
• Open DNN Box by Power Side-Channel Attack
• Analytical Moment Regularizer for Gaussian Robust Networks
• Beneficial perturbation network for continual learning