- TL;DR: Robustness performance of PGD trained models are sensitive to semantics-preserving transformation of image datasets, which implies the trickiness of evaluation of robust learning algorithms in practice.
- Abstract: Neural networks are vulnerable to small adversarial perturbations. While existing literature largely focused on the vulnerability of learned models, we demonstrate an intriguing phenomenon that adversarial robustness, unlike clean accuracy, is sensitive to the input data distribution. Even a semantics-preserving transformations on the input data distribution can cause a significantly different robustness for the adversarially trained model that is both trained and evaluated on the new distribution. We show this by constructing semantically- identical variants for MNIST and CIFAR10 respectively, and show that standardly trained models achieve similar clean accuracies on them, but adversarially trained models achieve significantly different robustness accuracies. This counter-intuitive phenomenon indicates that input data distribution alone can affect the adversarial robustness of trained neural networks, not necessarily the tasks themselves. Lastly, we discuss the practical implications on evaluating adversarial robustness, and make initial attempts to understand this complex phenomenon.