A3T: accuracy aware adversarial training

<div><p>Adversarial training has been empirically shown to be more prone to overfitting than standard training. The exact underlying reasons are still not fully understood. In this paper, we identify one cause of overfitting related to current practices of generating adversarial examples...

وصف كامل

محفوظ في:
التفاصيل البيبلوغرافية
المؤلف الرئيسي: Enes Altinisik (17725956) (author)
مؤلفون آخرون: Safa Messaoud (17725959) (author), Husrev Taha Sencar (17725962) (author), Sanjay Chawla (4254202) (author)
منشور في: 2023
الموضوعات:
الوسوم: إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
_version_ 1864513533725638656
author Enes Altinisik (17725956)
author2 Safa Messaoud (17725959)
Husrev Taha Sencar (17725962)
Sanjay Chawla (4254202)
author2_role author
author
author
author_facet Enes Altinisik (17725956)
Safa Messaoud (17725959)
Husrev Taha Sencar (17725962)
Sanjay Chawla (4254202)
author_role author
dc.creator.none.fl_str_mv Enes Altinisik (17725956)
Safa Messaoud (17725959)
Husrev Taha Sencar (17725962)
Sanjay Chawla (4254202)
dc.date.none.fl_str_mv 2023-07-14T03:00:00Z
dc.identifier.none.fl_str_mv 10.1007/s10994-023-06341-w
dc.relation.none.fl_str_mv https://figshare.com/articles/journal_contribution/A3T_accuracy_aware_adversarial_training/24934944
dc.rights.none.fl_str_mv CC BY 4.0
info:eu-repo/semantics/openAccess
dc.subject.none.fl_str_mv Information and computing sciences
Artificial intelligence
Software engineering
Adversarial training
Overftting in adversarial training
Accuracy aware adversarial training
dc.title.none.fl_str_mv A3T: accuracy aware adversarial training
dc.type.none.fl_str_mv Text
Journal contribution
info:eu-repo/semantics/publishedVersion
text
contribution to journal
description <div><p>Adversarial training has been empirically shown to be more prone to overfitting than standard training. The exact underlying reasons are still not fully understood. In this paper, we identify one cause of overfitting related to current practices of generating adversarial examples from misclassified samples. We show that, following current practice, adversarial examples from misclassified samples results in harder-to-classify samples than the original ones. This leads to a complex adjustment of the decision boundary during training and hence overfitting. To mitigate this issue, we propose A3T, an accuracy aware AT method that generate adversarial example differently for misclassified and correctly classified samples. We show that our approach achieves better generalization while maintaining comparable robustness to state-of-the-art AT methods on a wide range of computer vision, natural language processing, and tabular tasks.</p><p> </p></div><h2>Other Information</h2> <p> Published in: Machine Learning<br> License: <a href="https://creativecommons.org/licenses/by/4.0" target="_blank">https://creativecommons.org/licenses/by/4.0</a><br>See article on publisher's website: <a href="https://dx.doi.org/10.1007/s10994-023-06341-w" target="_blank">https://dx.doi.org/10.1007/s10994-023-06341-w</a></p>
eu_rights_str_mv openAccess
id Manara2_bdb0c1695bf565f7d37ea5a3ac86655d
identifier_str_mv 10.1007/s10994-023-06341-w
network_acronym_str Manara2
network_name_str Manara2
oai_identifier_str oai:figshare.com:article/24934944
publishDate 2023
repository.mail.fl_str_mv
repository.name.fl_str_mv
repository_id_str
rights_invalid_str_mv CC BY 4.0
spelling A3T: accuracy aware adversarial trainingEnes Altinisik (17725956)Safa Messaoud (17725959)Husrev Taha Sencar (17725962)Sanjay Chawla (4254202)Information and computing sciencesArtificial intelligenceSoftware engineeringAdversarial trainingOverftting in adversarial trainingAccuracy aware adversarial training<div><p>Adversarial training has been empirically shown to be more prone to overfitting than standard training. The exact underlying reasons are still not fully understood. In this paper, we identify one cause of overfitting related to current practices of generating adversarial examples from misclassified samples. We show that, following current practice, adversarial examples from misclassified samples results in harder-to-classify samples than the original ones. This leads to a complex adjustment of the decision boundary during training and hence overfitting. To mitigate this issue, we propose A3T, an accuracy aware AT method that generate adversarial example differently for misclassified and correctly classified samples. We show that our approach achieves better generalization while maintaining comparable robustness to state-of-the-art AT methods on a wide range of computer vision, natural language processing, and tabular tasks.</p><p> </p></div><h2>Other Information</h2> <p> Published in: Machine Learning<br> License: <a href="https://creativecommons.org/licenses/by/4.0" target="_blank">https://creativecommons.org/licenses/by/4.0</a><br>See article on publisher's website: <a href="https://dx.doi.org/10.1007/s10994-023-06341-w" target="_blank">https://dx.doi.org/10.1007/s10994-023-06341-w</a></p>2023-07-14T03:00:00ZTextJournal contributioninfo:eu-repo/semantics/publishedVersiontextcontribution to journal10.1007/s10994-023-06341-whttps://figshare.com/articles/journal_contribution/A3T_accuracy_aware_adversarial_training/24934944CC BY 4.0info:eu-repo/semantics/openAccessoai:figshare.com:article/249349442023-07-14T03:00:00Z
spellingShingle A3T: accuracy aware adversarial training
Enes Altinisik (17725956)
Information and computing sciences
Artificial intelligence
Software engineering
Adversarial training
Overftting in adversarial training
Accuracy aware adversarial training
status_str publishedVersion
title A3T: accuracy aware adversarial training
title_full A3T: accuracy aware adversarial training
title_fullStr A3T: accuracy aware adversarial training
title_full_unstemmed A3T: accuracy aware adversarial training
title_short A3T: accuracy aware adversarial training
title_sort A3T: accuracy aware adversarial training
topic Information and computing sciences
Artificial intelligence
Software engineering
Adversarial training
Overftting in adversarial training
Accuracy aware adversarial training