Titre :	Security of Machine Learning Systems against Adversarial Settings
Type de document :	texte imprimé
Auteurs :	Dounia Yessaad Cherif, Auteur ; Asma Bouhidel ; Yasmine Harbi, Directeur de thèse
Editeur :	Setif:UFA
Année de publication :	2024
Importance :	1 vol (64 f .)
Format :	29 cm
Langues :	Anglais (eng)
Catégories :	Thèses & Mémoires:Informatique
Mots-clés :	Machine Learning Adversarial attacks Convolutional neural networks, Healthcare applications Medical imaging
Index. décimale :	004 - Informatique
Résumé :	Machine learning (ML) has gained widespread adoption across various domains due to its significant advancements in recent years. It has shown remarkable efficacy in addressing complex tasks, often approaching or surpassing human-level capabilities. However, recent research has identified vulnerabilities such as adversarial attacks, which exploit weaknesses in models to induce erroneous predictions. These attacks pose substantial risks, particularly in critical domains such as autonomous vehicles and healthcare, underscoring the necessity for robust defenses to ensure the reliability and safety of machine learning systems. This thesis introduces a novel approach to bolster the resilience of convolutional neural network (CNN) models within the healthcare sector. Specifically, it focuses on enhancing performance in medical image tasks, including diagnosis, prognosis, and treatment planning. The study evaluates the robustness of CNN models against adversarial attacks and proposes effective defense strategies. Techniques such as adversarial training and integration of Generative Adversarial Networks (GANs) are employed to mitigate vulnerabilities, resulting in significant improvements in model resilience. The findings underscore the imperative of developing secure and dependable machine learning models in healthcare, paving the way for safer and more effective applications in clinical settings.
Note de contenu :	Sommaire Table of Contents ii List of Figures iv List of Tables vi Table of Abbreviations vii General Introduction 1 1 State-of-the-art 3 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2 Overview of Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2.1 ML Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.2.2 ML Workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.2.3 ML Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.2.4 ML Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.3 Adversarial Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.3.1 Origins of ML Attacks . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.3.2 Attack Surface in ML . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.3.3 AML Defense Methods . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.4 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2 Contribution 24 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.2 Problem Statement, Objectives, and Assumptions . . . . . . . . . . . . . 25 2.3 Proposed Model without Attacks . . . . . . . . . . . . . . . . . . . . . . . 25 2.3.1 Dataset Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.3.2 Dataset Splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.3.3 Dataset Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.3.4 Medical Image Classification . . . . . . . . . . . . . . . . . . . . . . 29 2.4 Proposed Model with Attacks . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.1 Adversarial Examples Generation . . . . . . . . . . . . . . . . . . . 31 2.5 Proposed Model with Attack and Defense . . . . . . . . . . . . . . . . . . 32 2.5.1 Adversarial Training . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.5.2 Generative Adversarial Network . . . . . . . . . . . . . . . . . . . 34 2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 3 Experiments and Results 38 3.1 Hardware and Software Environments . . . . . . . . . . . . . . . . . . . . 38 3.2 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.3 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.4.1 Proposed Model without Attacks . . . . . . . . . . . . . . . . . . . 42 3.4.2 Proposed Model with Attack . . . . . . . . . . . . . . . . . . . . . 47 3.4.3 Proposed Model with Attack and Defense . . . . . . . . . . . . . . 49 3.4.4 Comparison to Related Work . . . . . . . . . . . . . . . . . . . . . 56 3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Côte titre :	MAI/0897

Security of Machine Learning Systems against Adversarial Settings [texte imprimé] / Dounia Yessaad Cherif, Auteur ; Asma Bouhidel ; Yasmine Harbi, Directeur de thèse . - [S.l.] : Setif:UFA, 2024 . - 1 vol (64 f .) ; 29 cm.
Langues : Anglais (eng)

Catégories :	Thèses & Mémoires:Informatique
Mots-clés :	Machine Learning Adversarial attacks Convolutional neural networks, Healthcare applications Medical imaging
Index. décimale :	004 - Informatique
Résumé :	Machine learning (ML) has gained widespread adoption across various domains due to its significant advancements in recent years. It has shown remarkable efficacy in addressing complex tasks, often approaching or surpassing human-level capabilities. However, recent research has identified vulnerabilities such as adversarial attacks, which exploit weaknesses in models to induce erroneous predictions. These attacks pose substantial risks, particularly in critical domains such as autonomous vehicles and healthcare, underscoring the necessity for robust defenses to ensure the reliability and safety of machine learning systems. This thesis introduces a novel approach to bolster the resilience of convolutional neural network (CNN) models within the healthcare sector. Specifically, it focuses on enhancing performance in medical image tasks, including diagnosis, prognosis, and treatment planning. The study evaluates the robustness of CNN models against adversarial attacks and proposes effective defense strategies. Techniques such as adversarial training and integration of Generative Adversarial Networks (GANs) are employed to mitigate vulnerabilities, resulting in significant improvements in model resilience. The findings underscore the imperative of developing secure and dependable machine learning models in healthcare, paving the way for safer and more effective applications in clinical settings.
Note de contenu :	Sommaire Table of Contents ii List of Figures iv List of Tables vi Table of Abbreviations vii General Introduction 1 1 State-of-the-art 3 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2 Overview of Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2.1 ML Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.2.2 ML Workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.2.3 ML Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.2.4 ML Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.3 Adversarial Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.3.1 Origins of ML Attacks . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.3.2 Attack Surface in ML . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.3.3 AML Defense Methods . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.4 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2 Contribution 24 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.2 Problem Statement, Objectives, and Assumptions . . . . . . . . . . . . . 25 2.3 Proposed Model without Attacks . . . . . . . . . . . . . . . . . . . . . . . 25 2.3.1 Dataset Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.3.2 Dataset Splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.3.3 Dataset Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.3.4 Medical Image Classification . . . . . . . . . . . . . . . . . . . . . . 29 2.4 Proposed Model with Attacks . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.1 Adversarial Examples Generation . . . . . . . . . . . . . . . . . . . 31 2.5 Proposed Model with Attack and Defense . . . . . . . . . . . . . . . . . . 32 2.5.1 Adversarial Training . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.5.2 Generative Adversarial Network . . . . . . . . . . . . . . . . . . . 34 2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 3 Experiments and Results 38 3.1 Hardware and Software Environments . . . . . . . . . . . . . . . . . . . . 38 3.2 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.3 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.4.1 Proposed Model without Attacks . . . . . . . . . . . . . . . . . . . 42 3.4.2 Proposed Model with Attack . . . . . . . . . . . . . . . . . . . . . 47 3.4.3 Proposed Model with Attack and Defense . . . . . . . . . . . . . . 49 3.4.4 Comparison to Related Work . . . . . . . . . . . . . . . . . . . . . 56 3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Côte titre :	MAI/0897