|
|
|
|
LEADER |
01739 am a22002173u 4500 |
001 |
130090 |
042 |
|
|
|a dc
|
100 |
1 |
0 |
|a Tsipras, Dimitris
|e author
|
100 |
1 |
0 |
|a Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
|e contributor
|
700 |
1 |
0 |
|a Santurkar, Shibani
|q (Shibani Vinay)
|e author
|
700 |
1 |
0 |
|a Engstrom, Logan G.
|e author
|
700 |
1 |
0 |
|a Turner, Alexander M.
|e author
|
700 |
1 |
0 |
|a Madry, Aleksander
|e author
|
245 |
0 |
0 |
|a Robustness may be at odds with accuracy
|
260 |
|
|
|b ICLR,
|c 2021-03-05T14:59:33Z.
|
856 |
|
|
|z Get fulltext
|u https://hdl.handle.net/1721.1/130090
|
520 |
|
|
|a We show that there exists 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 even in a fairly simple and natural setting. These findings also corroborate a similar phenomenon observed in practice. 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 features learned by robust models tend to align better with salient data characteristics and human perception.
|
520 |
|
|
|a National Science Foundation (U.S.) (Grants S-1447786, IIS-1607189, CCF-1563880,CCF-1553428)
|
546 |
|
|
|a en
|
655 |
7 |
|
|a Article
|
773 |
|
|
|t 7th International Conference on Learning Representations, ICLR 2019
|