# Robustness May Be at Odds with Accuracy

Source: [Robustness May Be at Odds with Accuracy](<https://www.sota2.com/research/paper/68fa665c1e209131cea0fcea>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-25T21:48:29.028Z

Authors: Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, Aleksander Madry

Published: 2018-05-30T18:00:32.000Z

## Abstract

We show that there may exist an inherent tension between the goal of adversarial robustness and that of standard generalization. Specifically, training robust models may not only be more resource-consuming, but also lead to a reduction of standard accuracy. We demonstrate that this trade-off between the standard accuracy of a model and its robustness to adversarial perturbations provably exists in a fairly simple and natural setting. These findings also corroborate a similar phenomenon observed empirically in more complex settings. Further, we argue that this phenomenon is a consequence of robust classifiers learning fundamentally different feature representations than standard classifiers. These differences, in particular, seem to result in unexpected benefits: the representations learned by robust models tend to align better with salient data characteristics and human perception.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Image Classification on Caltech101 (PGD-l2 attack, ε=0.5, test)](<https://www.sota2.com/research/sota/image-classification-on-caltech101-pgd-l2-attack-0-5-test.md>): Image Classification; Caltech101 PGD-l2 attack, ε=0.5 (test); Robust Accuracy: 88.67.

- [Image Classification on DTD PGD-l2 attack, ε=0.5 (test)](<https://www.sota2.com/research/sota/image-classification-on-dtd-pgd-l2-attack-0-5-test.md>): Image Classification; DTD PGD-l2 attack, ε=0.5 (test); Robust Accuracy: 64.75.

- [Image Classification on Pets PGD-l2 attack, ε=0.5 (test)](<https://www.sota2.com/research/sota/image-classification-on-pets-pgd-l2-attack-0-5-test.md>): Image Classification; Pets PGD-l2 attack, ε=0.5 (test); Robust Accuracy: 79.53.

- [Image Classification on Caltech101 (PGD-l\_inf attack, ε=4/255, test)](<https://www.sota2.com/research/sota/image-classification-on-caltech101-pgd-l_inf-attack-4-255-test.md>): Image Classification; Caltech101 (PGD-l∞ attack, ε=4/255) (test); Robust Acc: 79.87.

- [Image Classification on DTD PGD-l∞ attack, ε=4/255 (test)](<https://www.sota2.com/research/sota/image-classification-on-dtd-pgd-l-attack-4-255-test.md>): Image Classification; DTD PGD-l∞ attack, ε=4/255 (test); Robust Accuracy: 48.95.

- [Image Classification on Pets PGD-l∞ attack, ε=4/255 (test)](<https://www.sota2.com/research/sota/image-classification-on-pets-pgd-l-attack-4-255-test.md>): Image Classification; Pets PGD-l∞ attack, ε=4/255 (test); Robust Accuracy: 65.6.

- [Image Classification on CIFAR10 (test) (Robustness & Clean Accuracy)](<https://www.sota2.com/research/sota/image-classification-on-cifar10-test-robustness-and-clean-accuracy.md>): Image Classification; CIFAR10 (test); Robust Accuracy: 90.74.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/1805.12152>)

[PDF](<https://arxiv.org/pdf/1805.12152.pdf>)
