{"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/improving-the-generalization-of-adversarial","title":"Improving the Generalization of Adversarial Training with Domain Adaptation","arxiv_id":"1810.00740","date":"2018-10-01","proceeding":"ICLR 2019 5","authors":["Chuanbiao Song","Kun He","Li-Wei Wang","John E. Hopcroft"],"abstract":"By injecting adversarial examples into training data, adversarial training is\npromising for improving the robustness of deep learning models. However, most\nexisting adversarial training approaches are based on a specific type of\nadversarial attack. It may not provide sufficiently representative samples from\nthe adversarial domain, leading to a weak generalization ability on adversarial\nexamples from other attacks. Moreover, during the adversarial training,\nadversarial perturbations on inputs are usually crafted by fast single-step\nadversaries so as to scale to large datasets. This work is mainly focused on\nthe adversarial training yet efficient FGSM adversary. In this scenario, it is\ndifficult to train a model with great generalization due to the lack of\nrepresentative adversarial samples, aka the samples are unable to accurately\nreflect the adversarial domain. To alleviate this problem, we propose a novel\nAdversarial Training with Domain Adaptation (ATDA) method. Our intuition is to\nregard the adversarial training on FGSM adversary as a domain adaption task\nwith limited number of target domain samples. The main idea is to learn a\nrepresentation that is semantically meaningful and domain invariant on the\nclean domain as well as the adversarial domain. Empirical evaluations on\nFashion-MNIST, SVHN, CIFAR-10 and CIFAR-100 demonstrate that ATDA can greatly\nimprove the generalization of adversarial training and the smoothness of the\nlearned models, and outperforms state-of-the-art methods on standard benchmark\ndatasets. To show the transfer ability of our method, we also extend ATDA to\nthe adversarial training on iterative attacks such as PGD-Adversial Training\n(PAT) and the defense performance is improved considerably.","url_abs":"http://arxiv.org/abs/1810.00740v7","url_pdf":"http://arxiv.org/pdf/1810.00740v7.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":"improving-the-generalization-of-adversarial","repo_url":"https://github.com/JHL-HUST/ATDA","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"tf","reach":{"status":"ok"}},{"paper_slug":"improving-the-generalization-of-adversarial","repo_url":"https://github.com/cxmscb/ATDA","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"tf","reach":{"status":"ok"}}],"tasks":[{"task_slug":"adversarial-attack","task_name":"Adversarial Attack"},{"task_slug":"domain-adaptation","task_name":"Domain Adaptation"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1810.00740","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}