{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/adversarial-logit-pairing","title":"Adversarial Logit Pairing","arxiv_id":"1803.06373","date":"2018-03-16","proceeding":"NeurIPS 2018","authors":["Harini Kannan","Alexey Kurakin","Ian Goodfellow"],"abstract":"In this paper, we develop improved techniques for defending against\nadversarial examples at scale. First, we implement the state of the art version\nof adversarial training at unprecedented scale on ImageNet and investigate\nwhether it remains effective in this setting - an important open scientific\nquestion (Athalye et al., 2018). Next, we introduce enhanced defenses using a\ntechnique we call logit pairing, a method that encourages logits for pairs of\nexamples to be similar. When applied to clean examples and their adversarial\ncounterparts, logit pairing improves accuracy on adversarial examples over\nvanilla adversarial training; we also find that logit pairing on clean examples\nonly is competitive with adversarial training in terms of accuracy on two\ndatasets. Finally, we show that adversarial logit pairing achieves the state of\nthe art defense on ImageNet against PGD white box attacks, with an accuracy\nimprovement from 1.5% to 27.9%. Adversarial logit pairing also successfully\ndamages the current state of the art defense against black box attacks on\nImageNet (Tramer et al., 2018), dropping its accuracy from 66.6% to 47.1%. With\nthis new accuracy drop, adversarial logit pairing ties with Tramer et al.(2018)\nfor the state of the art on black box attacks on ImageNet.","url_abs":"http://arxiv.org/abs/1803.06373v1","url_pdf":"http://arxiv.org/pdf/1803.06373v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"adversarial-logit-pairing","repo_url":"https://github.com/arobey1/advbench","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null},{"paper_slug":"adversarial-logit-pairing","repo_url":"https://github.com/labsix/adversarial-logit-pairing-analysis","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"Apache-2.0"}},{"paper_slug":"adversarial-logit-pairing","repo_url":"https://github.com/tensorflow/models/tree/master/research/adversarial_logit_pairing","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":null},{"paper_slug":"adversarial-logit-pairing","repo_url":"https://github.com/uds-lsv/evaluating-logit-pairing-methods","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"Apache-2.0"}}],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1803.06373","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}