AI, Machine Learning and Deep Learning: A Security Perspective

✍ Scribed by Fei Hu, Xiali Hei

• Publisher: CRC Press
• Year: 2023
• Tongue: English
• Leaves: 347
• Category: Library

✦ Synopsis

Today, Artificial Intelligence (AI) and Machine Learning/ Deep Learning (ML/DL) have become the hottest areas in information technology. In our society, many intelligent devices rely on AI/ML/DL algorithms/tools for smart operations. Although AI/ML/DL algorithms and tools have been used in many internet applications and electronic devices, they are also vulnerable to various attacks and threats. AI parameters may be distorted by the internal attacker; the DL input samples may be polluted by adversaries; the ML model may be misled by changing the classification boundary, among many other attacks and threats. Such attacks can make AI products dangerous to use.

While this discussion focuses on security issues in AI/ML/DL-based systems (i.e., securing the intelligent systems themselves), AI/ML/DL models and algorithms can actually also be used for cyber security (i.e., the use of AI to achieve security).

Since AI/ML/DL security is a newly emergent field, many researchers and industry professionals cannot yet obtain a detailed, comprehensive understanding of this area. This book aims to provide a complete picture of the challenges and solutions to related security issues in various applications. It explains how different attacks can occur in advanced AI tools and the challenges of overcoming those attacks. Then, the book describes many sets of promising solutions to achieve AI security and privacy. The features of this book have seven aspects

This is the first book to explain various practical attacks and countermeasures to AI systems.
Both quantitative math models and practical security implementations are provided.
It covers both "securing the AI system itself" and "using AI to achieve security".
It covers all the advanced AI attacks and threats with detailed attack models.
It provides multiple solution spaces to the security and privacy issues in AI tools.
The differences among ML and DL security and privacy issues are explained.
Many practical security applications are covered.

✦ Table of Contents

Cover
Half Title
Title Page
Copyright Page
Table of Contents
Preface
About the Editors
Contributors
Part I Secure AI/ML Systems: Attack Models
1 Machine Learning Attack Models
1.1 Introduction
1.2 Background
1.2.1 Notation
1.2.2 Support Vector Machines
1.2.3 Neural Networks
1.3 White-Box Adversarial Attacks
1.3.1 L-BGFS Attack
1.3.2 Fast Gradient Sign Method
1.3.3 Basic Iterative Method
1.3.4 DeepFool
1.3.5 Fast Adaptive Boundary Attack
1.3.6 Carlini and Wagner’s Attack
1.3.7 Shadow Attack
1.3.8 Wasserstein Attack
1.4 Black-Box Adversarial Attacks
1.4.1 Transfer Attack
1.4.2 Score-Based Black-Box Attacks
ZOO Attack
Square Attack
1.4.3 Decision-Based Attack
Boundary Attack
HopSkipJump Attack
Spatial Transformation Attack
1.5 Data Poisoning Attacks
1.5.1 Label Flipping Attacks
1.5.2 Clean Label Data Poisoning Attack
Feature Collision Attack
Convex Polytope Attack and Bullseye Polytope Attack
1.5.3 Backdoor Attack
1.6 Conclusions
Acknowledgment
Note
References
2 Adversarial Machine Learning: A New Threat Paradigm for Next-Generation Wireless Communications
2.1 Introduction
2.1.1 Scope and Background
2.2 Adversarial Machine Learning
2.3 Challenges and Gaps
2.3.1 Development Environment
2.3.2 Training and Test Datasets
2.3.3 Repeatability, Hyperparameter Optimization, and Explainability
2.3.4 Embedded Implementation
2.4 Conclusions and Recommendations
References
3 Threat of Adversarial Attacks to Deep Learning: A Survey
3.1 Introduction
3.2 Categories of Attacks
3.2.1 White-Box Attacks
FGSM-based Method
JSMA-based Method
3.2.2 Black-Box Attacks
Mobility-based Approach
Gradient Estimation-Based Approach
3.3 Attacks Overview
3.3.1 Attacks On Computer-Vision-Based Applications
3.3.2 Attacks On Natural Language Processing Applications
3.3.3 Attacks On Data Poisoning Applications
3.4 Specific Attacks In The Real World
3.4.1 Attacks On Natural Language Processing
3.4.2 Attacks Using Data Poisoning
3.5 Discussions and Open Issues
3.6 Conclusions
References
4 Attack Models for Collaborative Deep Learning
4.1 Introduction
4.2 Background
4.2.1 Deep Learning (DL)
Convolution Neural Network
4.2.2 Collaborative Deep Learning (CDL)
Architecture
Collaborative Deep Learning Workflow
4.2.3 Deep Learning Security and Collaborative Deep Learning Security
4.3 Auror: An Automated Defense
4.3.1 Problem Setting
4.3.2 Threat Model
Targeted Poisoning Attacks
4.3.3 AUROR Defense
4.3.4 Evaluation
4.4 A New CDL Attack: Gan Attack
4.4.1 Generative Adversarial Network (GAN)
4.4.2 GAN Attack
Main Protocol
4.4.3 Experiment Setups
Dataset
System Architecture
Hyperparameter Setup
4.4.4 Evaluation
4.5 Defend Against Gan Attack In IoT
4.5.1 Threat Model
4.5.2 Defense System
4.5.3 Main Protocols
4.5.4 Evaluation
4.6 Conclusions
Acknowledgment
References
5 Attacks On Deep Reinforcement Learning Systems: A Tutorial
5.1 Introduction
5.2 Characterizing Attacks on DRL Systems
5.3 Adversarial Attacks
5.4 Policy Induction Attacks
5.5 Conclusions and Future Directions
References
6 Trust and Security of Deep Reinforcement Learning
6.1 Introduction
6.2 Deep Reinforcement Learning Overview
6.2.1 Markov Decision Process
6.2.2 Value-Based Methods
V-value Function
Q-value Function
Advantage Function
Bellman Equation
6.2.3 Policy-Based Methods
6.2.4 Actor–Critic Methods
6.2.5 Deep Reinforcement Learning
6.3 The Most Recent Reviews
6.3.1 Adversarial Attack On Machine Learning
6.3.1.1 Evasion Attack
6.3.1.2 Poisoning Attack
6.3.2 Adversarial Attack On Deep Learning
6.3.2.1 Evasion Attack
6.3.2.2 Poisoning Attack
6.3.3 Adversarial Deep Reinforcement Learning
6.4 Attacks On DRL Systems
6.4.1 Attacks On Environment
6.4.2 Attacks On States
6.4.3 Attacks On Policy Function
6.4.4 Attacks On Reward Function
6.5 Defenses Against DRL System Attacks
6.5.1 Adversarial Training
6.5.2 Robust Learning
6.5.3 Adversarial Detection
6.6 Robust DRL Systems
6.6.1 Secure Cloud Platform
6.6.2 Robust DRL Modules
6.7 A Scenario of Financial Stability
6.7.1 Automatic Algorithm Trading Systems
6.8 Conclusion and Future Work
References
7 IoT Threat Modeling Using Bayesian Networks
7.1 Background
7.2 Topics of Chapter
7.3 Scope
7.4 Cyber Security In IoT Networks
7.4.1 Smart Home
7.4.2 Attack Graphs
7.5 Modeling With Bayesian Networks
7.5.1 Graph Theory
7.5.2 Probabilities and Distributions
7.5.3 Bayesian Networks
7.5.4 Parameter Learning
7.5.5 Inference
7.6 Model Implementation
7.6.1 Network Structure
7.6.2 Attack Simulation
Selection Probabilities
Vulnerability Probabilities Based On CVSS Scores
Attack Simulation Algorithm
7.6.3 Network Parametrization
7.6.4 Results
7.7 Conclusions and Future Work
References
Part II Secure AI/ML Systems: Defenses
8 Survey of Machine Learning Defense Strategies
8.1 Introduction
8.2 Security Threats
8.3 Honeypot Defense
8.4 Poisoned Data Defense
8.5 Mixup Inference Against Adversarial Attacks
8.6 Cyber-Physical Techniques
8.7 Information Fusion Defense
8.8 Conclusions and Future Directions
References
9 Defenses Against Deep Learning Attacks
9.1 Introduction
9.2 Categories of Defenses
9.2.1 Modified Training Or Modified Input
Data Preprocessing
Data Augmentation
9.2.2 Modifying Networks Architecture
Network Distillation
Model Regularization
9.2.3 Network Add-On
Defense Against Universal Perturbations
MegNet Model
9.4 Discussions and Open Issues
9.5 Conclusions
References
10 Defensive Schemes for Cyber Security of Deep Reinforcement Learning
10.1 Introduction
10.2 Background
10.2.1 Model-Free RL
10.2.2 Deep Reinforcement Learning
10.2.3 Security of DRL
10.3 Certificated Verification For Adversarial Examples
10.3.1 Robustness Certification
10.3.2 System Architecture
10.3.3 Experimental Results
10.4 Robustness On Adversarial State Observations
10.4.1 State-Adversarial DRL for Deterministic Policies: DDPG
10.4.2 State-Adversarial DRL for Q-Learning: DQN
10.4.3 Experimental Results
10.5 Conclusion And Challenges
Acknowledgment
References
11 Adversarial Attacks On Machine Learning Models in Cyber- Physical Systems
11.1 Introduction
11.2 Support Vector Machine (SVM) Under Evasion Attacks
11.2.1 Adversary Model
11.2.2 Attack Scenarios
11.2.3 Attack Strategy
11.3 SVM Under Causality Availability Attack
11.4 Adversarial Label Contamination on SVM
11.4.1 Random Label Flips
11.4.2 Adversarial Label Flips
11.5 Conclusions
References
12 Federated Learning and Blockchain:: An Opportunity for Artificial Intelligence With Data Regulation
12.1 Introduction
12.2 Data Security And Federated Learning
12.3 Federated Learning Context
12.3.1 Type of Federation
12.3.1.1 Model-Centric Federated Learning
12.3.1.2 Data-Centric Federated Learning
12.3.2 Techniques
12.3.2.1 Horizontal Federated Learning
12.3.2.2 Vertical Federated Learning
12.4 Challenges
12.4.1 Trade-Off Between Efficiency and Privacy
12.4.2 Communication Bottlenecks
12.4.3 Poisoning
12.5 Opportunities
12.5.1 Leveraging Blockchain
12.6 Use Case: Leveraging Privacy, Integrity, And Availability For Data-Centric Federated Learning Using A Blockchain-Based Approach
12.6.1 Results
12.7 Conclusion
References
Part III Using AI/ML Algorithms for Cyber Security
13 Using Machine Learning for Cyber Security: Overview
13.1 Introduction
13.2 Is Artificial Intelligence Enough To Stop Cyber Crime?
13.3 Corporations’ Use Of Machine Learning To Strengthen Their Cyber Security Systems
13.4 Cyber Attack/Cyber Security Threats And Attacks
13.4.1 Malware
13.4.2 Data Breach
13.4.3 Structured Query Language Injection (SQL-I)
13.4.4 Cross-Site Scripting (XSS)
13.4.5 Denial-Of-Service (DOS) Attack
13.4.6 Insider Threats
13.4.7 Birthday Attack
13.4.8 Network Intrusions
13.4.9 Impersonation Attacks
13.4.10 DDoS Attacks Detection On Online Systems
13.5 Different Machine Learning Techniques In Cyber Security
13.5.1 Support Vector Machine (SVM)
13.5.2 K-Nearest Neighbor (KNN)
13.5.3 Naïve Bayes
13.5.4 Decision Tree
13.5.5 Random Forest (RF)
13.5.6 Multilayer Perceptron (MLP)
13.6 Application Of Machine Learning
13.6.1 ML in Aviation Industry
13.6.2 Cyber ML Under Cyber Security Monitoring
13.6.3 Battery Energy Storage System (BESS) Cyber Attack Mitigation
13.6.4 Energy-Based Cyber Attack Detection in Large-Scale Smart Grids
13.6.5 IDS for Internet of Vehicles (IoV)
13.7 Deep Learning Techniques In Cyber Security
13.7.1 Deep Auto-Encoder
13.7.2 Convolutional Neural Networks (CNN)
13.7.3 Recurrent Neural Networks (RNNs)
13.7.4 Deep Neural Networks (DNNs)
13.7.5 Generative Adversarial Networks (GANs)
13.7.6 Restricted Boltzmann Machine (RBM)
13.7.7 Deep Belief Network (DBN)
13.8 Applications Of Deep Learning In Cyber Security
13.8.1 Keystroke Analysis
13.8.2 Secure Communication in IoT
13.8.3 Botnet Detection
13.8.4 Intrusion Detection and Prevention Systems (IDS/IPS)
13.8.5 Malware Detection in Android
13.8.6 Cyber Security Datasets
13.8.7 Evaluation Metrics
13.9 CONCLUSION
References
14 Performance of Machine Learning and Big Data Analytics Paradigms in Cyber Security
14.1 Introduction
14.1.1 Background On Cyber Security and Machine Learning
14.1.2 Background Perspectives to Big Data Analytics and Cyber Security
14.1.3 Supervised Learning Algorithms
14.1.4 Statement of the Problem
14.1.5 Purpose of Study
14.1.6 Research Objectives
14.1.7 Research Questions
14.2 LITERATURE REVIEW
14.2.1 Overview
14.2.2 Classical Machine Learning (CML)
14.2.2.1 Logistic Regression (LR)
14.2.2.2 Naïve Bayes (NB)
14.2.2.3 Decision Tree (DT)
14.2.2.4 K-Nearest Neighbor (KNN)
14.2.2.5 AdaBoost (AB)
14.2.2.6 Random Forest (RF)
14.2.2.7 Support Vector Machine (SVM)
14.2.3 Modern Machine Learning
14.2.3.1 Deep Neural Network (DNN)
14.2.3.2 Future of AI in the Fight Against Cyber Crimes
14.2.4 Big Data Analytics and Cyber Security
14.2.4.1 Big Data Analytics Issues
14.2.4.2 Independent Variable: Big Data Analytics
14.2.4.3 Intermediating Variables
14.2.4.4 Conceptual Framework
14.2.4.5 Theoretical Framework
14.2.4.6 Big Data Analytics Application to Cyber Security
14.2.4.7 Big Data Analytics and Cyber Security Limitations
14.2.4.8 Limitations
14.2.5 Advances in Cloud Computing
14.2.5.1 Explaining Cloud Computing and How It Has Evolved to Date
14.2.6 Cloud Characteristics
14.2.7 Cloud Computing Service Models
14.2.7.1 Software as a Service (SaaS)
14.2.7.2 Platform as a Service (PaaS)
14.2.7.3 Infrastructure as a Service (IaaS)
14.2.8 Cloud Deployment Models
14.2.8.1 Private Cloud
14.2.8.2 Public Cloud
14.2.8.3 Hybrid Cloud
14.2.8.4 Community Cloud
14.2.8.5 Advantages and Disadvantages of Cloud Computing
14.2.8.6 Six Main Characteristics of Cloud Computing and How They Are Leveraged
14.2.8.7 Some Advantages of Network Function Virtualization
14.2.8.8 Virtualization and Containerization Compared and Contrasted
14.3 Research Methodology
14.3.1 Presentation of the Methodology
14.3.1.1 Research Approach and Philosophy
14.3.1.2 Research Design and Methods
14.3.2 Population and Sampling
14.3.2.1 Population
14.3.2.2 Sample
14.3.3 Sources and Types of Data
14.3.4 Model for Analysis
14.3.4.1 Big Data
14.3.4.2 Big Data Analytics
14.3.4.3 Insights for Action
14.3.4.4 Predictive Analytics
14.3.5 Validity and Reliability
14.3.6 Summary of Research Methodology
14.3.7 Possible Outcomes
14.4 Analysis And Research Outcomes
14.4.1 Overview
14.4.2 Support Vector Machine
14.4.3 KNN Algorithm
14.4.4 Multilinear Discriminant Analysis (LDA)
14.4.5 Random Forest Classifier
14.4.6 Variable Importance
14.4.7 Model Results
14.4.8 Classification and Regression Trees (CART)
14.4.9 Support Vector Machine
14.4.10 Linear Discriminant Algorithm
14.4.11 K-Nearest Neighbor
14.4.12 Random Forest
14.4.13 Challenges and Future Direction
14.4.13.1 Model 1: Experimental/Prototype Model
14.4.13.2 Model 2: Cloud Computing/Outsourcing
14.4.13.3 Application of Big Data Analytics Models in Cyber Security
14.4.13.4 Summary of Analysis
14.5 Conclusion
References
15 Using ML and DL Algorithms for Intrusion Detection in the Industrial Internet of Things
15.1 Introduction
15.2 IDS Applications
15.2.1 Random Forest Classifier
15.2.2 Pearson Correlation Coefficient
15.2.3 Related Works
15.3 Use Of ML And DL Algorithms In IIOT Applications
15.4 Practical Application of ML Algorithms In IIOT
15.4.1 Results
15.5 Conclusion
References
Part IV Applications
16 On Detecting Interest Flooding Attacks in Named Data Networking (NDN)–based IoT Searches
16.1 Introduction
16.2 PRELIMINARIES
16.2.1 Named Data Networking (NDN)
16.2.2 Internet of Things Search Engine (IoTSE)
16.2.3 Machine Learning (ML)
16.3 Machine Learning Assisted For NDN-Based IFA Detection In IOTSE
16.3.1 Attack Model
16.3.2 Attack Scale
16.3.3 Attack Scenarios
16.3.4 Machine Learning (ML) Detection Models
16.4 Performance Evaluation
16.4.1 Methodology
16.4.2 IFA Performance
16.4.2.1 Simple Tree Topology (Small Scale)
16.4.2.2 Rocketfuel ISP Like Topology (Large Scale)
16.4.3 Data Processing for Detection
16.4.4 Detection Results
16.4.4.1 ML Detection Performance in Simple Tree Topology
16.4.4.2 ML Detection in Rocketful ISP Topology
16.5 Discussion
16.6 Related Works
16.7 Final Remarks
Acknowledgment
References
17 Attack On Fraud Detection Systems in Online Banking Using Generative Adversarial Networks
17.1 Introduction
17.1.1 Problem of Fraud Detection in Banking
17.1.2 Fraud Detection and Prevention System
17.2 Experiment Description
17.2.1 Research Goal
17.2.2 Empirical Data
17.2.3 Attack Scenario
17.3 Generator And Discrimination Model
17.3.1 Model Construction
17.3.1.1 Imitation Fraud Detection System Model
17.3.1.2 Generator Models
17.3.2 Evaluation of Models
17.4 Final Conclusions and Recommendations
Notes
References
18 Artificial Intelligence-Assisted Security Analysis of Smart Healthcare Systems
18.1 Introduction
18.2 Smart Healthcare System (SHS)
18.2.1 Formal Modeling of SHS
18.2.2 Machine Learning (ML)–based Patient Status Classification Module (PSCM) in SHS
18.2.2.1 Decision Tree (DT)
18.2.2.2 Logistic Regression (LR)
18.2.2.3 Neural Network (NN)
18.2.3 Hyperparameter Optimization of PSCM in SHS
18.2.3.1 Whale Optimization (WO)
18.2.3.2 Grey Wolf Optimization (GWO)
18.2.3.3 Firefly Optimization (FO)
18.2.3.4 Evaluation Results
18.3 Formal Attack Modeling of SHS
18.3.1 Attacks in SHS
18.3.2 Attacker’s Knowledge
18.3.3 Attacker’s Capability
18.3.4 Attacker’s Accessibility
18.3.5 Attacker’s Goal
18.4 Anomaly Detection Models (ADMS) In SHS
18.4.1 ML-Based Anomaly Detection Model (ADM) in SHS
18.4.1.1 Density-Based Spatial Clustering of Applications With Noise (DBSCAN)
18.4.1.2 K-Means
18.4.1.3 One-Class SVM (OCSVM)
18.4.1.4 Autoencoder (AE)
18.4.2 Ensemble-Based ADMs in SHS
18.4.2.1 Data Collection and Preprocessing
18.4.2.2 Model Training
18.4.2.3 Threshold Calculation
18.4.2.4 Anomaly Detection
18.4.2.5 Example Case Studies
18.4.2.6 Evaluation Result
18.4.2.7 Hyperparameter Optimization of ADMs in SHS
18.5 Formal Attack Analysis of Smart Healthcare Systems
18.5.1 Example Case Studies
18.5.2 Performance With Respect to Attacker Capability
18.5.3 Frequency of Sensors in the Attack Vectors
18.5.4 Scalability Analysis
18.6 Resiliency Analysis Of Smart Healthcare System
18.7 Conclusion and Future Works
References
19 A User-Centric Focus for Detecting Phishing Emails
19.1 Introduction
19.2 Background and Related Work
19.2.1 Behavioral Models Related to Phishing Susceptibility
19.2.2 User-Centric Antiphishing Measures
19.2.3 Technical Antiphishing Measures
19.2.4 Research Gap
19.3 The Dataset
19.4 Understanding The Decision Behavior Of Machine Learning Models
19.4.1 Interpreter for Machine Learning Algorithms
19.4.2 Local Interpretable Model-Agnostic Explanations (LIME)
19.4.3 Anchor Explanations
19.4.3.1 Share of Emails in the Data for Which the Rule Holds
19.5 Designing the Artifact
19.5.1 Background
19.5.2 Identifying Suspected Phishing Attempts
19.5.3 Cues in Phishing Emails
19.5.4 Extracting Cues
19.5.5 Examples of the Application of XAI for Extracting Cues and Phrases
19.6 Conclusion and Future Works
19.6.1 Completion of the Artifact
Notes
References