<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Publishing DTD v1.3 20210610//EN" "JATS-journalpublishing1-3.dtd">
<article article-type="research-article" dtd-version="1.3" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xml:lang="ru"><front><journal-meta><journal-id journal-id-type="publisher-id">ntv</journal-id><journal-title-group><journal-title xml:lang="ru">Научно-технический вестник информационных технологий, механики и оптики</journal-title><trans-title-group xml:lang="en"><trans-title>Scientific and Technical Journal of Information Technologies, Mechanics and Optics</trans-title></trans-title-group></journal-title-group><issn pub-type="ppub">2226-1494</issn><issn pub-type="epub">2500-0373</issn><publisher><publisher-name>Университет ИТМО</publisher-name></publisher></journal-meta><article-meta><article-id pub-id-type="doi">10.17586/2226-1494-2024-24-2-256-266</article-id><article-id custom-type="elpub" pub-id-type="custom">ntv-96</article-id><article-categories><subj-group subj-group-type="heading"><subject>Research Article</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="ru"><subject>ИСКУССТВЕННЫЙ ИНТЕЛЛЕКТ И КОГНИТИВНЫЕ ИНФОРМАЦИОННЫЕ ТЕХНОЛОГИИ</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="en"><subject>ARTIFICIAL INTELLIGENCE AND COGNITIVE INFORMATION TECHNOLOGIES</subject></subj-group></article-categories><title-group><article-title>Новый метод противодействия состязательным атакам уклонения на информационные системы, основанные на искусственном интеллекте</article-title><trans-title-group xml:lang="en"><trans-title>A new method for countering evasion adversarial attacks on information systems based on artificial intelligence</trans-title></trans-title-group></title-group><contrib-group><contrib contrib-type="author" corresp="yes"><contrib-id contrib-id-type="orcid">https://orcid.org/0000-0001-6691-6167</contrib-id><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Воробьева</surname><given-names>А. А.</given-names></name><name name-style="western" xml:lang="en"><surname>Vorobeva</surname><given-names>A. A.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Воробьева Алиса Андреевна — кандидат технических наук, доцент </p><p> Санкт-Петербург, 197101 </p><p> sc 57191359167 </p></bio><bio xml:lang="en"><p> Alisa A. Vorobeva — PhD, Associate Professor </p><p> Saint Petersburg, 197101 </p><p> sc 57191359167 </p></bio><email xlink:type="simple">vorobeva@itmo.ru</email><xref ref-type="aff" rid="aff-1"/></contrib><contrib contrib-type="author" corresp="yes"><contrib-id contrib-id-type="orcid">https://orcid.org/0009-0006-2179-6847</contrib-id><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Матузко</surname><given-names>М. А,</given-names></name><name name-style="western" xml:lang="en"><surname>Matuzko</surname><given-names>M. A.</given-names></name></name-alternatives><bio xml:lang="ru"><p> Матузко Максим Александрович — инженер </p><p> Санкт-Петербург, 197101 </p></bio><bio xml:lang="en"><p>Maxim A. Matuzko — Engineer </p><p> Saint Petersburg, 197101 </p></bio><email xlink:type="simple">mxmmtzk@gmail.com</email><xref ref-type="aff" rid="aff-1"/></contrib><contrib contrib-type="author" corresp="yes"><contrib-id contrib-id-type="orcid">https://orcid.org/0000-0002-3008-3789</contrib-id><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Сивков</surname><given-names>Д. И,</given-names></name><name name-style="western" xml:lang="en"><surname>Sivkov</surname><given-names>D. I.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Сивков Дмитрий Игоревич — инженер </p><p> Санкт-Петербург, 197101 </p></bio><bio xml:lang="en"><p> Dmitry I. Sivkov — Engineer </p><p> Saint Petersburg, 197101 </p></bio><email xlink:type="simple">485280@mail.ru</email><xref ref-type="aff" rid="aff-1"/></contrib><contrib contrib-type="author" corresp="yes"><contrib-id contrib-id-type="orcid">https://orcid.org/0009-0004-8635-9432</contrib-id><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Сафиуллин</surname><given-names>Р. И.</given-names></name><name name-style="western" xml:lang="en"><surname>Safiullin</surname><given-names>R. I.</given-names></name></name-alternatives><bio xml:lang="ru"><p> Сафиуллин Роман Ильшатович — инженер</p><p> Санкт-Петербург, 197101 </p></bio><bio xml:lang="en"><p> Roman I. Safiullin — Engineer </p><p> Saint Petersburg, 197101 </p></bio><email xlink:type="simple">romsaaf@mail.ru</email><xref ref-type="aff" rid="aff-1"/></contrib><contrib contrib-type="author" corresp="yes"><contrib-id contrib-id-type="orcid">https://orcid.org/0000-0002-2287-4310</contrib-id><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Менщиков</surname><given-names>А. А,</given-names></name><name name-style="western" xml:lang="en"><surname>Menshchikov</surname><given-names>A. A.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Менщиков Александр Алексеевич — кандидат технических наук, доцент</p><p> Санкт-Петербург, 197101 </p><p> sc 57196052901 </p></bio><bio xml:lang="en"><p> Alexander A. Menshchikov — PhD, Associate Professor</p><p> Saint Petersburg, 197101 </p></bio><email xlink:type="simple">menshikov@itmo.ru</email><xref ref-type="aff" rid="aff-1"/></contrib></contrib-group><aff-alternatives id="aff-1"><aff xml:lang="ru"><institution>Университет ИТМО</institution><country>Россия</country></aff><aff xml:lang="en"><institution>ITMO University</institution><country>Russian Federation</country></aff></aff-alternatives><pub-date pub-type="collection"><year>2024</year></pub-date><pub-date pub-type="epub"><day>12</day><month>12</month><year>2024</year></pub-date><volume>24</volume><issue>2</issue><fpage>256</fpage><lpage>266</lpage><permissions><copyright-statement>Copyright &amp;#x00A9; Воробьева А.А., Матузко М.А., Сивков Д.И., Сафиуллин Р.И., Менщиков А.А., 2024</copyright-statement><copyright-year>2024</copyright-year><copyright-holder xml:lang="ru">Воробьева А.А., Матузко М.А., Сивков Д.И., Сафиуллин Р.И., Менщиков А.А.</copyright-holder><copyright-holder xml:lang="en">Vorobeva A.A., Matuzko M.A., Sivkov D.I., Safiullin R.I., Menshchikov A.A.</copyright-holder><license xml:lang="ru" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>Данная работа распространяется под лицензией Creative Commons Attribution 4.0.</license-p></license><license xml:lang="en" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>This work is licensed under a Creative Commons Attribution 4.0 License.</license-p></license></permissions><self-uri xlink:href="https://ntv.elpub.ru/jour/article/view/96">https://ntv.elpub.ru/jour/article/view/96</self-uri><abstract><p>Введение. Современные технологии искусственного интеллекта находят применение в различных областях науки и повседневной жизни. Повсеместное внедрение систем, основанных на методах искусственного интеллекта, выявило проблему их уязвимости перед состязательными атаками, включающими методы обмана искусственной нейронной сети и нарушения ее работы. В работе основное внимание уделено защите моделей распознавания изображений от состязательных атак уклонения, признанных в настоящее время наиболее опасными. При таких атаках создаются состязательные данные, содержащие незначительные искажения относительно исходных, и происходит отправка их на обученную модель с целью изменения ее «ответа» на вариант, необходимый злоумышленнику. Искажения могут включать добавление шума или изменение нескольких пикселов изображения. Рассмотрены наиболее актуальные подходы к созданию состязательных данных: метод быстрого градиента (Fast Gradient Sign Method, FGSM), метод квадрата (Square Method, SQ), метод прогнозируемого градиентного спуска (Predicted Gradient Descent, PGD), базовый итеративный метод (Basic Iterative Method, BIM), метод Карлини и Вагнера (Carlini-Wagner, CW), метод карт значимости Якобиана (Jacobian Saliency Map Attack, JSMA). Исследованы современные методы противодействия атакам уклонения, основанные на модификации модели — состязательное обучение и предварительная обработка поступающих данных: пространственное сглаживание, сжатие признаков, JPEG-сжатие, минимизация общей дисперсии, оборонительная дистилляция. Эти методы эффективны только против определенных видов атак. На сегодняшний день ни один метод противодействия не может быть применен в качестве универсального решения. Метод. Предложен новый метод, сочетающий состязательное обучение с предварительной обработкой изображений. Состязательное обучение выполнено на основе состязательных данных, создаваемых с распространенных атак, что позволяет эффективно им противодействовать. Предварительная обработка изображений предназначена для противодействия атакам, которые не учитывались при состязательном обучении, что дает возможность защитить систему от атак новых типов. Обработка осуществлена методом JPEG-сжатия и сжатия признаков для уменьшения влияния состязательных искажений и более эффективного противодействия всем видам рассмотренных атак. Основные результаты. Проведена оценка показателей качества распознавания изображений на основе искусственной нейронной сети. Экспериментальные данные включали оригинальные и измененные изображения, созданные с использованием методов атак типов FGSM, PGD, BIM, SQ, CW, JSMA. При этом состязательное обучение модели в экспериментах выполнено на данных, содержащих состязательные примеры только для методов атак FGSM, PGD, BIM. Набор данных, использованный в экспериментах, являлся сбалансированным. Оценена средняя точность распознавания изображений, в условиях отправки на модель изображений, созданных с использованием указанных видов атак. Сделаны выводы, что состязательное обучение эффективно только для противодействия атакам, которые использовались во время обучения модели, а методы предварительной обработки поступающих данных эффективны только против более простых атак. Средняя точность распознавания в случае применения разработанного метода составила 0,94, что существенно выше рассмотренных методов противодействия атакам. Показано, что точность без применения методов противодействия составляет величину около 0,19, а при состязательном обучении — 0,79, пространственном сглаживании — 0,58, сжатии признаков — 0,88, JPEG-сжатии — 0,37, минимизации общей дисперсии — 0,58, оборонительной дистилляция — 0,44. При этом точность распознавания при атаках FGM, PGD, BIM, SQ, CW, JSMA составила соответственно 0,99, 0,99, 0,98, 0,98, 0,99, 0,73. Разработанный метод представляет более универсальное решение по противодействию всем видам атак, а также достаточно эффективно работает при противодействии сложным состязательным атакам, таким как атаки CW и JSMA. Обсуждение. Разработанный метод позволяет повысить точность распознавания с применением машинного обучения при атаках уклонения и, в отличие от состязательного обучения, повышает точность распознавания на состязательных данных, создаваемых с применением атак, не использованных при обучении. Полученные результаты полезны исследователям и специалистам в области машинного обучения.</p></abstract><trans-abstract xml:lang="en"><p>Modern artificial intelligence (AI) technologies are being used in a variety of fields, from science to everyday life. However, the widespread use of AI-based systems has highlighted a problem with their vulnerability to adversarial attacks. These attacks include methods of fooling or misleading an artificial neural network, disrupting its operations, and causing it to make incorrect predictions. This study focuses on protecting image recognition models against adversarial evasion attacks which have been recognized as the most challenging and dangerous. In these attacks, adversaries create adversarial data that contains minor perturbations compared to the original image, and then send it to a trained model in an attempt to change its response to the desired outcome. These distortions can involve adding noise or even changing a few pixels. In this paper, we consider the most relevant methods for generating adversarial data: the Fast Gradient Sign Method (FGSM), the Square Method (SQ), the predicted gradient descent method (PGD), the Basic Iterative Method (BIM), the Carlini-Wagner method (CW) and Jacobian Saliency Map Attack (JSMA). We also study modern techniques for defending against evasion attacks through model modification, such as adversarial training and pre-processing of incoming data, including spatial smoothing, feature squeezing, jpeg compression, minimizing total variance, and defensive distillation. While these methods are effective against certain types of attacks, to date, there is no single method that can be used as a universal defense. Instead, we propose a new method that combines adversarial learning with image pre-processing. We suggest that adversarial training should be performed on adversarial samples generated from common attack methods which can then be effectively defended against. The image preprocessing aims to counter attacks that were not considered during adversarial training. This allows to protect the system from new types of attacks. It is proposed to use jpeg compression and feature squeezing on the pre-processing stage. This reduces the impact of adversarial perturbations and effectively counteracts all types of considered attacks. The evaluation of image recognition model (based on convolutional neural network) performance metrics based was conducted. The experimental data included original images and adversarial images created using attack FGSM, PGD, BIM, SQ, CW, and JSMA methods. At the same time, adversarial training of the model was performed in experiments on data containing only adversarial examples for the FGSM, PGD, and BIM attack methods. Dataset used in experiments was balanced. The average accuracy of image recognition was estimated with crafted adversarial imaged datasets. It was concluded that adversarial training is effective only in countering attacks that were used during model training, while methods of pre-processing incoming data are effective only against more simple attacks. The average recognition accuracy using the developed method was 0.94, significantly higher than those considered methods for countering attacks. It has been shown that the accuracy without using any counteraction methods is approximately 0.19, while with adversarial learning it is 0.79. Spatial smoothing provides an accuracy of 0.58, and feature squeezing results in an accuracy of 0.88. Jpeg compression provides an accuracy of 0.37, total variance minimization — 0.58 and defensive distillation — 0.44. At the same time, image recognition accuracy provided by developed method for FGSM, PGD, BIM, SQ, CW, and JSMA attacks is 0.99, 0.99, 0.98, 0.98, 0.99 and 0.73, respectively. The developed method is a more universal solution for countering all types of attacks and works quite effectively against complex adversarial attacks such as CW and JSMA. The developed method makes it possible to increase accuracy of image recognition model for adversarial images. Unlike adversarial learning, it also increases recognition accuracy on adversarial data generated using attacks not used on training stage. The results are useful for researchers and practitioners in the field of machine learning.</p></trans-abstract><kwd-group xml:lang="ru"><kwd>методы машинного обучения</kwd><kwd>состязательные атаки</kwd><kwd>защитные механизмы</kwd><kwd>информационные системы на базе искусственного интеллекта</kwd><kwd>состязательное обучение</kwd></kwd-group><kwd-group xml:lang="en"><kwd>machine learning methods</kwd><kwd>adversarial attacks</kwd><kwd>defense mechanisms</kwd><kwd>AI-based information systems</kwd><kwd>adversarial learning</kwd></kwd-group></article-meta></front><back><ref-list><title>References</title><ref id="cit1"><label>1</label><citation-alternatives><mixed-citation xml:lang="ru">Szegedy C., Zaremba W., Sutskever I., Bruna J., Erhan D., Goodfellow I., Fergus R. Intriguing properties of neural networks // arXiv. 2013. arXiv:1312.6199. https://doi.org/10.48550/arXiv.1312.6199</mixed-citation><mixed-citation xml:lang="en">Szegedy C., Zaremba W., Sutskever I., Bruna J., Erhan D., Goodfellow I., Fergus R. Intriguing properties of neural networks // arXiv. 2013. arXiv:1312.6199. https://doi.org/10.48550/arXiv.1312.6199</mixed-citation></citation-alternatives></ref><ref id="cit2"><label>2</label><citation-alternatives><mixed-citation xml:lang="ru">Tabassi E., Burns K.J., Hadjimichael M., Molina-Markham A.D., Sexton J.T. A taxonomy and terminology of adversarial machine learning: NIST IR. 2019. P. 1–29.</mixed-citation><mixed-citation xml:lang="en">Tabassi E., Burns K.J., Hadjimichael M., Molina-Markham A.D., Sexton J.T. A taxonomy and terminology of adversarial machine learning: NIST IR. 2019. P. 1–29.</mixed-citation></citation-alternatives></ref><ref id="cit3"><label>3</label><citation-alternatives><mixed-citation xml:lang="ru">Goodfellow I.J., Shlens J., Szegedy C. Explaining and harnessing adversarial examples // arXiv. 2015. arXiv:1412.6572. https://doi.org/10.48550/arXiv.1412.6572</mixed-citation><mixed-citation xml:lang="en">Goodfellow I.J., Shlens J., Szegedy C. Explaining and harnessing adversarial examples // arXiv. 2015. arXiv:1412.6572. https://doi.org/10.48550/arXiv.1412.6572</mixed-citation></citation-alternatives></ref><ref id="cit4"><label>4</label><citation-alternatives><mixed-citation xml:lang="ru">Carlini N., Mishra P., Vaidya T., Zhang Y., Sherr M., Shields C., Wagner D., Zhou W. Hidden voice commands // Proc. of the 25th USENIX Security Symposium. 2016. P. 513–530.</mixed-citation><mixed-citation xml:lang="en">Carlini N., Mishra P., Vaidya T., Zhang Y., Sherr M., Shields C., Wagner D., Zhou W. Hidden voice commands // Proc. of the 25th USENIX Security Symposium. 2016. P. 513–530.</mixed-citation></citation-alternatives></ref><ref id="cit5"><label>5</label><citation-alternatives><mixed-citation xml:lang="ru">Zhang G., Yan C., Ji X., Zhang T., Zhang T., Xu W. Dolphinattack: Inaudible voice commands // Proc. of the 2017 ACM SIGSAC Conference on Computer and Communications Security. 2017. P. 103–117. https://doi.org/10.1145/3133956.3134052</mixed-citation><mixed-citation xml:lang="en">Zhang G., Yan C., Ji X., Zhang T., Zhang T., Xu W. Dolphinattack: Inaudible voice commands // Proc. of the 2017 ACM SIGSAC Conference on Computer and Communications Security. 2017. P. 103–117. https://doi.org/10.1145/3133956.3134052</mixed-citation></citation-alternatives></ref><ref id="cit6"><label>6</label><citation-alternatives><mixed-citation xml:lang="ru">Kurakin A., Goodfellow I.J., Bengio S. Adversarial machine learning at scale // International Conference on Learning Representations (ICLR). 2017.</mixed-citation><mixed-citation xml:lang="en">Kurakin A., Goodfellow I.J., Bengio S. Adversarial machine learning at scale // International Conference on Learning Representations (ICLR). 2017.</mixed-citation></citation-alternatives></ref><ref id="cit7"><label>7</label><citation-alternatives><mixed-citation xml:lang="ru">Li X., Zhu D. Robust detection of adversarial attacks on medical images // Proc. of the 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI). 2020. P. 1154–1158. https://doi.org/10.1109/isbi45749.2020.9098628</mixed-citation><mixed-citation xml:lang="en">Li X., Zhu D. Robust detection of adversarial attacks on medical images // Proc. of the 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI). 2020. P. 1154–1158. https://doi.org/10.1109/isbi45749.2020.9098628</mixed-citation></citation-alternatives></ref><ref id="cit8"><label>8</label><citation-alternatives><mixed-citation xml:lang="ru">Imam N.H., Vassilakis V.G. A survey of attacks against twitter spam detectors in an adversarial environment // Robotics. 2019. V. 8. N 3. P. 50. https://doi.org/10.3390/robotics8030050</mixed-citation><mixed-citation xml:lang="en">Imam N.H., Vassilakis V.G. A survey of attacks against twitter spam detectors in an adversarial environment // Robotics. 2019. V. 8. N 3. P. 50. https://doi.org/10.3390/robotics8030050</mixed-citation></citation-alternatives></ref><ref id="cit9"><label>9</label><citation-alternatives><mixed-citation xml:lang="ru">Andriushchenko M., Croce F., Flammarion N., Hein M. Square attack: a query-efficient black-box adversarial attack via random search // Lecture Notes in Computer Science. 2020. V. 12368. P. 484– 501. https://doi.org/10.1007/978-3-030-58592-1_29</mixed-citation><mixed-citation xml:lang="en">Andriushchenko M., Croce F., Flammarion N., Hein M. Square attack: a query-efficient black-box adversarial attack via random search // Lecture Notes in Computer Science. 2020. V. 12368. P. 484– 501. https://doi.org/10.1007/978-3-030-58592-1_29</mixed-citation></citation-alternatives></ref><ref id="cit10"><label>10</label><citation-alternatives><mixed-citation xml:lang="ru">Deng Y., Karam L.J. Universal adversarial attack via enhanced projected gradient descent // Proc. of the 2020 IEEE International Conference on Image Processing (ICIP). 2020. P. 1241–1245. https://doi.org/10.1109/icip40778.2020.9191288</mixed-citation><mixed-citation xml:lang="en">Deng Y., Karam L.J. Universal adversarial attack via enhanced projected gradient descent // Proc. of the 2020 IEEE International Conference on Image Processing (ICIP). 2020. P. 1241–1245. https://doi.org/10.1109/icip40778.2020.9191288</mixed-citation></citation-alternatives></ref><ref id="cit11"><label>11</label><citation-alternatives><mixed-citation xml:lang="ru">Madry A., Makelov A., Schmidt L., Tsipras D., Vladu A. Towards deep learning models resistant to adversarial attacks // International Conference on Learning Representations (ICLR). 2018. 12. Kurakin A., Goodfellow I.J., Bengio S. Adversarial examples in the physical world // Artificial Intelligence Safety and Security. 2018. P. 99–112. https://doi.org/10.1201/9781351251389-8</mixed-citation><mixed-citation xml:lang="en">Madry A., Makelov A., Schmidt L., Tsipras D., Vladu A. Towards deep learning models resistant to adversarial attacks // International Conference on Learning Representations (ICLR). 2018. 12. Kurakin A., Goodfellow I.J., Bengio S. Adversarial examples in the physical world // Artificial Intelligence Safety and Security. 2018. P. 99–112. https://doi.org/10.1201/9781351251389-8</mixed-citation></citation-alternatives></ref><ref id="cit12"><label>12</label><citation-alternatives><mixed-citation xml:lang="ru">Carlini N., Wagner D. Towards evaluating the robustness of neural networks // Proc. of the 2017 IEEE Symposium on Security and Privacy (SP). 2017. P. 39–57. https://doi.org/10.1109/sp.2017.49</mixed-citation><mixed-citation xml:lang="en">Carlini N., Wagner D. Towards evaluating the robustness of neural networks // Proc. of the 2017 IEEE Symposium on Security and Privacy (SP). 2017. P. 39–57. https://doi.org/10.1109/sp.2017.49</mixed-citation></citation-alternatives></ref><ref id="cit13"><label>13</label><citation-alternatives><mixed-citation xml:lang="ru">Papernot N., McDaniel P., Jha S., Fredrikson M., Celik Z.B., Swami A. The limitations of deep learning in adversarial settings // Proc. of the 2016 IEEE European Symposium on Security and Privacy (EuroS&amp;P). 2016. P. 372–387. https://doi.org/10.1109/eurosp.2016.36</mixed-citation><mixed-citation xml:lang="en">Papernot N., McDaniel P., Jha S., Fredrikson M., Celik Z.B., Swami A. The limitations of deep learning in adversarial settings // Proc. of the 2016 IEEE European Symposium on Security and Privacy (EuroS&amp;P). 2016. P. 372–387. https://doi.org/10.1109/eurosp.2016.36</mixed-citation></citation-alternatives></ref><ref id="cit14"><label>14</label><citation-alternatives><mixed-citation xml:lang="ru">Lowd D., Meek C. Adversarial learning // Proc. of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining. 2005. P. 641–647. https://doi.org/10.1145/1081870.1081950</mixed-citation><mixed-citation xml:lang="en">Lowd D., Meek C. Adversarial learning // Proc. of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining. 2005. P. 641–647. https://doi.org/10.1145/1081870.1081950</mixed-citation></citation-alternatives></ref><ref id="cit15"><label>15</label><citation-alternatives><mixed-citation xml:lang="ru">Das N., Shanbhogue M., Chen S.-T., Hohman F., Chen L., Kounavis M.E., Chau D.H. Keeping the bad guys out: Protecting and vaccinating deep learning with JPEG compression // arXiv. 2017. arXiv:1705.02900. https://doi.org/10.48550/arXiv.1705.02900</mixed-citation><mixed-citation xml:lang="en">Das N., Shanbhogue M., Chen S.-T., Hohman F., Chen L., Kounavis M.E., Chau D.H. Keeping the bad guys out: Protecting and vaccinating deep learning with JPEG compression // arXiv. 2017. arXiv:1705.02900. https://doi.org/10.48550/arXiv.1705.02900</mixed-citation></citation-alternatives></ref><ref id="cit16"><label>16</label><citation-alternatives><mixed-citation xml:lang="ru">Guo C., Rana M., Cisse M., van der Maaten L. Countering adversarial images using input transformations // International Conference on Learning Representations (ICLR). 2018.</mixed-citation><mixed-citation xml:lang="en">Guo C., Rana M., Cisse M., van der Maaten L. Countering adversarial images using input transformations // International Conference on Learning Representations (ICLR). 2018.</mixed-citation></citation-alternatives></ref><ref id="cit17"><label>17</label><citation-alternatives><mixed-citation xml:lang="ru">Xu W., Evans D., Qi Y. Feature squeezing: detecting adversarial examples in deep neural networks // Proc. of the 2018 Network and Distributed System Security Symposium. 2018. https://doi.org/10.14722/ndss.2018.23198</mixed-citation><mixed-citation xml:lang="en">Xu W., Evans D., Qi Y. Feature squeezing: detecting adversarial examples in deep neural networks // Proc. of the 2018 Network and Distributed System Security Symposium. 2018. https://doi.org/10.14722/ndss.2018.23198</mixed-citation></citation-alternatives></ref><ref id="cit18"><label>18</label><citation-alternatives><mixed-citation xml:lang="ru">Papernot N., McDaniel P., Wu X., Jha S., Swami A. Distillation as a defense to adversarial perturbations against deep neural networks // Proc. of the 2016 IEEE Symposium on Security and Privacy (SP). 2016. P. 582–597. https://doi.org/10.1109/sp.2016.41</mixed-citation><mixed-citation xml:lang="en">Papernot N., McDaniel P., Wu X., Jha S., Swami A. Distillation as a defense to adversarial perturbations against deep neural networks // Proc. of the 2016 IEEE Symposium on Security and Privacy (SP). 2016. P. 582–597. https://doi.org/10.1109/sp.2016.41</mixed-citation></citation-alternatives></ref></ref-list><fn-group><fn fn-type="conflict"><p>The authors declare that there are no conflicts of interest present.</p></fn></fn-group></back></article>
