https:\/\/telecom-paris.zoom.us\/j\/98578578906?pwd=BGRr0jBb7hNZ1pn7kijxKIiAlNwseg.1<\/a><\/p>\n\n\n\nMembres du jury :<\/strong><\/p>\n\n\n\nM. Mihai MITREA<\/strong>, Professeur, T\u00e9l\u00e9com SudParis, FRANCE – Directeur de these
M. Stefano TUBARO<\/strong>, Professeur, Polytechnic University of Milan, ITALIE – Rapporteur
M. Azzedine BEGHDADI<\/strong>, Professeur des universit\u00e9s, Sorbonne Paris North University, FRANCE – Rapporteur
M. Michel KIEFFER<\/strong>, Professeur, Paris-Saclay University, FRANCE – Examinateur
Mme Rossella CANCELLIERE<\/strong>, Ma\u00eetre de conf\u00e9rences, University of Turin, ITALIE – Examinateur
Mme Benedetta TONDI<\/strong>, Ma\u00eetre de conf\u00e9rences, University of Siena, ITALIE – Examinateur
Mme Vicky KALOGEITON<\/strong>, Ma\u00eetre de conf\u00e9rences, Ecole Polytechnique, FRANCE – Examinateur<\/p>\n\n\n\n\u00ab Tatouage Robuste pour les R\u00e9seaux de Neurones dans les Applications Multim\u00e9dias \u00bb<\/h2>\n\n\n\npr\u00e9sent\u00e9 par Monsieur Carl DE SOUSA TRIAS<\/h2>\n\n\n\n
R\u00e9sum\u00e9 :<\/strong><\/p>\n\n\n\nLa th\u00e8se traite du suivi et de l’authentification des r\u00e9seaux de neurones (NN) par le tatouage (watermarking) robuste pour des t\u00e2ches applicatives li\u00e9es au traitement multim\u00e9dia (classification, segmentation, d\u00e9tection d\u2019objet et compression pour les images et la vid\u00e9o). Elle explore \u00e9galement la possibilit\u00e9 d’\u00e9tablir les bases communes permettant la sp\u00e9cification des normes ouvertes pour le tatouage des NN. Ainsi, une analyse approfondie de l\u2019\u00e9tat de l\u2019art est r\u00e9alis\u00e9e, aboutissant \u00e0 des nouvelles taxonomies compl\u00e9mentaires et \u00e0 un environnement de test de r\u00e9f\u00e9rence. Puis, l\u2019\u00e9tat de l\u2019art du tatouage de r\u00e9seaux de neurones est \u00e9tendu par quatre contributions m\u00e9thodologique. D\u2019abord, la nature des attaques (et contre-attaques) g\u00e9om\u00e9triques est formalis\u00e9e. Puis, des synergies m\u00e9thodologiques et compl\u00e9mentaire avec des outils d\u2019optimisation de l’entra\u00eenement des NN (e.g. SAM \u2013 Sharpness Aware Minimization) sont \u00e9tablies. Le r\u00e9sultat principal est une m\u00e9thode de white-box watermarking (d\u00e9tection avec acc\u00e8s au mod\u00e8le tatou\u00e9) et montre comment la complexit\u00e9 de calcul peut \u00eatre r\u00e9duite tout en pr\u00e9servant les propri\u00e9t\u00e9s de la m\u00e9thode de tatouage. Finalement, la relation \u00e9troite entre l\u2019entrainement sur la t\u00e2che et l\u2019insertion de la marque est \u00e9tudi\u00e9e par deux m\u00e9thodes compl\u00e9mentaires. D\u2019un c\u00f4t\u00e9, une m\u00e9thode de black-box watermarking (avec acc\u00e8s seulement aux inf\u00e9rences du mod\u00e8le tatou\u00e9) est d\u00e9finie. Reposant uniquement sur un terme de r\u00e9gularisation durant l\u2019apprentissage du mod\u00e8le, cette m\u00e9thode permet d\u2019\u00e9viter les limitations actuelles li\u00e9es \u00e0 l’utilisation de jeux de donn\u00e9es additionnels ou \u00e0 la n\u00e9cessit\u00e9 de mod\u00e8les NN suppl\u00e9mentaires con\u00e7us pour le tatouage. D\u2019un autre c\u00f4t\u00e9, une nouvelle m\u00e9thode de white-box watermarking d\u00e9montre la possibilit\u00e9 de d\u00e9coupler, de mani\u00e8re m\u00e9thodologique et computationnelle, l\u2019entrainement et l\u2019insertion de la marque. Parall\u00e8lement, afin d’assurer une \u00e9valuation objective des r\u00e9sultats dans le domaine du tatouage des NN et de favoriser les \u00e9changes entre les mondes acad\u00e9mique et industriel, la th\u00e8se a contribu\u00e9 activement aux travaux de normalisation relatifs au Neural Network Watermarking, aboutissant \u00e0 deux standards MPAI et IEEE.<\/p>\n\n\n\n
Abstract :<\/strong><\/p>\n\n\n\nThe present thesis deals with neural network (NN) tracking and authentication by means of robust and imperceptible watermarking. It focuses on application tasks related to multimedia processing (image\/video classification, segmentation, object detection, and compression), and investigates the possibility of building a common ground for the NN watermarking. To this end, a thorough analysis of the state-of-the-art studies is first carried out, thus bringing to light new taxonomies and providing open-source reference testbeds for NN watermarking assessment. Then, the state-of-the-art watermarking landscape is extended by four types of methodological contributions. First, the nature of geometric attacks is stated and the corresponding counterattack is formalized. Then, methodological synergies and complementarities with tools in NN training optimization (e.g. SAM \u2013 Sharpness Aware Minimization) are established. The main result is represented by a white-box watermarking method and shows how computational complexity can be traded for optimization solutions while preserving the watermarking properties. Finally, the inner relationship between the task-oriented learning and the mark insertion is investigated by two complementary approaches. On the one hand, the methodological grounds for black-box watermarking insertion solely based on a regularization term considered during the task-oriented training of the initial NN model are established; this way, current-day limitations related to the usage of modified datasets or to the need of additional NN models specifically designed for watermarking purposes are avoided. On the other hand, the possibility of logically and computationally decoupling the watermark insertion from the task-oriented learning process is demonstrated. In parallel, in order to jointly ensure an objective evaluation of the results in the field and to accelerate the academia-industry exchanges, the thesis contributed to intensive standardization activities related to Neural Network Watermarking, resulting in two MPAI and IEEE standards.<\/p>\n\n\n\n
Cette th\u00e8se a \u00e9t\u00e9 dirig\u00e9e par Prof. M. Mitrea<\/strong>, Prof. M. Cagnazzo<\/strong>, MdC A. Fiandrotti<\/strong> et co-encadr\u00e9 avec MdC S. Chaudhuri<\/strong> et MdC E. Tartaglione<\/strong>.<\/p>\n","protected":false},"excerpt":{"rendered":"L’Ecole doctorale : Ecole Doctorale de l’Institut Polytechnique de Paris et le Laboratoire de recherche SAMOVAR – Services r\u00e9partis, Architectures, Mod\u00e9lisation, Validation, Administration des R\u00e9seaux pr\u00e9sentent l\u2019AVIS DE SOUTENANCE de Monsieur Carl DE SOUSA TRIAS Autoris\u00e9 \u00e0 pr\u00e9senter ses travaux en vue de l\u2019obtention du Doctorat de l’Institut Polytechnique de Paris, pr\u00e9par\u00e9 \u00e0 T\u00e9l\u00e9com SudParis […]<\/p>\n","protected":false},"author":4,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ocean_post_layout":"","ocean_both_sidebars_style":"","ocean_both_sidebars_content_width":0,"ocean_both_sidebars_sidebars_width":0,"ocean_sidebar":"","ocean_second_sidebar":"","ocean_disable_margins":"enable","ocean_add_body_class":"","ocean_shortcode_before_top_bar":"","ocean_shortcode_after_top_bar":"","ocean_shortcode_before_header":"","ocean_shortcode_after_header":"","ocean_has_shortcode":"","ocean_shortcode_after_title":"","ocean_shortcode_before_footer_widgets":"","ocean_shortcode_after_footer_widgets":"","ocean_shortcode_before_footer_bottom":"","ocean_shortcode_after_footer_bottom":"","ocean_display_top_bar":"default","ocean_display_header":"default","ocean_header_style":"","ocean_center_header_left_menu":"","ocean_custom_header_template":"","ocean_custom_logo":0,"ocean_custom_retina_logo":0,"ocean_custom_logo_max_width":0,"ocean_custom_logo_tablet_max_width":0,"ocean_custom_logo_mobile_max_width":0,"ocean_custom_logo_max_height":0,"ocean_custom_logo_tablet_max_height":0,"ocean_custom_logo_mobile_max_height":0,"ocean_header_custom_menu":"","ocean_menu_typo_font_family":"","ocean_menu_typo_font_subset":"","ocean_menu_typo_font_size":0,"ocean_menu_typo_font_size_tablet":0,"ocean_menu_typo_font_size_mobile":0,"ocean_menu_typo_font_size_unit":"px","ocean_menu_typo_font_weight":"","ocean_menu_typo_font_weight_tablet":"","ocean_menu_typo_font_weight_mobile":"","ocean_menu_typo_transform":"","ocean_menu_typo_transform_tablet":"","ocean_menu_typo_transform_mobile":"","ocean_menu_typo_line_height":0,"ocean_menu_typo_line_height_tablet":0,"ocean_menu_typo_line_height_mobile":0,"ocean_menu_typo_line_height_unit":"","ocean_menu_typo_spacing":0,"ocean_menu_typo_spacing_tablet":0,"ocean_menu_typo_spacing_mobile":0,"ocean_menu_typo_spacing_unit":"","ocean_menu_link_color":"","ocean_menu_link_color_hover":"","ocean_menu_link_color_active":"","ocean_menu_link_background":"","ocean_menu_link_hover_background":"","ocean_menu_link_active_background":"","ocean_menu_social_links_bg":"","ocean_menu_social_hover_links_bg":"","ocean_menu_social_links_color":"","ocean_menu_social_hover_links_color":"","ocean_disable_title":"default","ocean_disable_heading":"default","ocean_post_title":"","ocean_post_subheading":"","ocean_post_title_style":"","ocean_post_title_background_color":"","ocean_post_title_background":0,"ocean_post_title_bg_image_position":"","ocean_post_title_bg_image_attachment":"","ocean_post_title_bg_image_repeat":"","ocean_post_title_bg_image_size":"","ocean_post_title_height":0,"ocean_post_title_bg_overlay":0.5,"ocean_post_title_bg_overlay_color":"","ocean_disable_breadcrumbs":"default","ocean_breadcrumbs_color":"","ocean_breadcrumbs_separator_color":"","ocean_breadcrumbs_links_color":"","ocean_breadcrumbs_links_hover_color":"","ocean_display_footer_widgets":"default","ocean_display_footer_bottom":"default","ocean_custom_footer_template":"","ocean_post_oembed":"","ocean_post_self_hosted_media":"","ocean_post_video_embed":"","ocean_link_format":"","ocean_link_format_target":"self","ocean_quote_format":"","ocean_quote_format_link":"post","ocean_gallery_link_images":"on","ocean_gallery_id":[],"footnotes":""},"categories":[286,169],"tags":[],"class_list":["post-6860","post","type-post","status-publish","format-standard","hentry","category-fractualites-ennews-fr","category-seminaires-armedia","entry"],"_links":{"self":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/6860","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/comments?post=6860"}],"version-history":[{"count":1,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/6860\/revisions"}],"predecessor-version":[{"id":6861,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/posts\/6860\/revisions\/6861"}],"wp:attachment":[{"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/media?parent=6860"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/categories?post=6860"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/samovar.telecom-sudparis.eu\/index.php\/wp-json\/wp\/v2\/tags?post=6860"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}