Tight certificates of adversarial robustness for randomly smoothed classifiers
Strong theoretical guarantees of robustness can be given for ensembles of classifiers generated by input randomization. Specifically, an `2 bounded adversary cannot alter the ensemble prediction generated by an additive isotropic Gaussian noise, where the radius for the adversary depends on both the...
| Main Authors: | , , |
|---|---|
| Other Authors: | |
| Format: | Article |
| Language: | English |
| Published: |
2021
|
| Online Access: | https://hdl.handle.net/1721.1/129439 |
| _version_ | 1826200792171806720 |
|---|---|
| author | Lee, Guang-He Yuan, Yang Jaakkola, Tommi S |
| author2 | Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory |
| author_facet | Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory Lee, Guang-He Yuan, Yang Jaakkola, Tommi S |
| author_sort | Lee, Guang-He |
| collection | MIT |
| description | Strong theoretical guarantees of robustness can be given for ensembles of classifiers generated by input randomization. Specifically, an `2 bounded adversary cannot alter the ensemble prediction generated by an additive isotropic Gaussian noise, where the radius for the adversary depends on both the variance of the distribution as well as the ensemble margin at the point of interest. We build on and considerably expand this work across broad classes of distributions. In particular, we offer adversarial robustness guarantees and associated algorithms for the discrete case where the adversary is `0 bounded. Moreover, we exemplify how the guarantees can be tightened with specific assumptions about the function class of the classifier such as a decision tree. We empirically illustrate these results with and without functional restrictions across image and molecule datasets. |
| first_indexed | 2024-09-23T11:41:29Z |
| format | Article |
| id | mit-1721.1/129439 |
| institution | Massachusetts Institute of Technology |
| language | English |
| last_indexed | 2024-09-23T11:41:29Z |
| publishDate | 2021 |
| record_format | dspace |
| spelling | mit-1721.1/1294392022-09-27T21:17:51Z Tight certificates of adversarial robustness for randomly smoothed classifiers Lee, Guang-He Yuan, Yang Jaakkola, Tommi S Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory Strong theoretical guarantees of robustness can be given for ensembles of classifiers generated by input randomization. Specifically, an `2 bounded adversary cannot alter the ensemble prediction generated by an additive isotropic Gaussian noise, where the radius for the adversary depends on both the variance of the distribution as well as the ensemble margin at the point of interest. We build on and considerably expand this work across broad classes of distributions. In particular, we offer adversarial robustness guarantees and associated algorithms for the discrete case where the adversary is `0 bounded. Moreover, we exemplify how the guarantees can be tightened with specific assumptions about the function class of the classifier such as a decision tree. We empirically illustrate these results with and without functional restrictions across image and molecule datasets. 2021-01-19T15:26:13Z 2021-01-19T15:26:13Z 2020-02 2019-06 2020-12-21T16:05:25Z Article http://purl.org/eprint/type/ConferencePaper 1049-5258 https://hdl.handle.net/1721.1/129439 Lee, Guang-He et al. “Tight certificates of adversarial robustness for randomly smoothed classifiers.”32nd Conference on Neural Information Processing Systems, December 2018, Montreal, Canada, Neural Information Processing Systems, 2018. © 2018 The Author(s) en https://papers.nips.cc/paper/2019/hash/fa2e8c4385712f9a1d24c363a2cbe5b8-Abstract.html 32nd Conference on Neural Information Processing Systems (NeurIPS 2018) Article is made available in accordance with the publisher's policy and may be subject to US copyright law. Please refer to the publisher's site for terms of use. application/pdf Neural Information Processing Systems (NIPS) |
| spellingShingle | Lee, Guang-He Yuan, Yang Jaakkola, Tommi S Tight certificates of adversarial robustness for randomly smoothed classifiers |
| title | Tight certificates of adversarial robustness for randomly smoothed classifiers |
| title_full | Tight certificates of adversarial robustness for randomly smoothed classifiers |
| title_fullStr | Tight certificates of adversarial robustness for randomly smoothed classifiers |
| title_full_unstemmed | Tight certificates of adversarial robustness for randomly smoothed classifiers |
| title_short | Tight certificates of adversarial robustness for randomly smoothed classifiers |
| title_sort | tight certificates of adversarial robustness for randomly smoothed classifiers |
| url | https://hdl.handle.net/1721.1/129439 |
| work_keys_str_mv | AT leeguanghe tightcertificatesofadversarialrobustnessforrandomlysmoothedclassifiers AT yuanyang tightcertificatesofadversarialrobustnessforrandomlysmoothedclassifiers AT jaakkolatommis tightcertificatesofadversarialrobustnessforrandomlysmoothedclassifiers |