A Discussion of 'Adversarial Examples Are Not Bugs, They Are Features': Two Examples of Useful, Non-Robust Features

25 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Shortcut learning in deep neural networks (2020, background): “If a successful adversarial attack [8] can change model predictions without changing semantic content, this is an indication that something akin to shortcut learning may be occurring [33, 97].”
A Self-supervised Approach for Adversarial Robustness (2020, background): “• Transferability of Natural vs. Robust Layers: In addition to style transfer hypothesis [5], we explore the connection between layers of VGG and adversarially trained models in the context of adversarial attacks: – Maximum Distortion of Non-Robust Features: Datasets containing natural images…”
Rethinking and Improving the Robustness of Image Style Transfer (2021, methodology): “To validate this conjecture, [27, 4] trained a ResNet with adversarial examples [10] to improve feature robustness.”