Autonomous Cyber Resilience

Charles A. Kamhoua, Ph.D., is a Researcher at the DEVCOM Army Research Laboratory Network Security Branch. Alexander Kott, Ph.D., is the Chief Scientist at the DEVCOM Army Research Laboratory. Quanyan Zhu, Ph.D., is an Associate Professor in the Department of Electrical and Computer Engineering at New York University. Nandi O. Leslie, Ph.D., is a Principal Technical Fellow in Raytheon Engineering at RTX.

• Preface xvEditor Biographies xvii1 Introduction 1Alexander Kott1.1 Cyber Resilience and Cybersecurity 11.2 Autonomy and Cyber Resilience 31.3 Autonomous Actions 41.4 Approaches to Implementing Autonomous Cyber Resilience 51.5 Benefits and Risks of Autonomous Cyber Resilience 81.6 The Preview of the Book 91.6.1 Part I: Foundations of Cyber Resilience 91.6.2 Part II: Resilient Machine Learning 101.6.3 Part III: Game Theory for Network Resilience 10References 11Part 1 Cyber Resilience 132 Game-theoretic Foundations for Cyber Resilience Against Deceptive Information Attacks in Intelligent Transportation Systems 15Ya-Ting Yang and Quanyan Zhu2.1 Introduction 152.1.1 Multi-domain Threats in ITS 172.1.2 Security Risk Assessment 182.1.3 Chapter Organization 192.2 Deceptive Information Attacks 192.2.1 Intra-vehicle Domain 202.2.2 Inter-vehicle Domain Attacks 222.2.3 Transportation System Domain 232.2.4 Human Aspects 272.3 Cross-layer Resilience 292.3.1 Holistic Framework for Resilience 302.3.2 Theoretical Foundations and Design Frameworks 312.3.3 Benefits of Game-theoretic and Learning-based Design Principles for Cyber Resilience 342.4 Case Study 352.4.1 Misinformation Attacks on Recommendation Systems 352.5 Conclusion and Discussion 432.5.1 Conclusion 432.5.2 Discussion 43References 443 CYBER-MIRA: Cyber Mission Impact Resilience Assessment Framework for Tactical Mission Systems 53Ashrith Reddy Thukkaraju, Han Jun Yoon, Shou Matsumoto, Jair Feldens Ferrari, Donghwan Lee, Myung Kil Ahn, Paulo Costa, and Jin-Hee Cho3.1 Introduction 533.1.1 Motivation and Challenges 543.1.2 Research Goal 553.1.3 Key Contributions 553.1.4 Structure of This Chapter 563.2 Related Work 563.2.1 Methodologies of CMIA 563.2.2 MIA Frameworks 573.2.3 Cyber Resilience Assessment 583.3 System Model 593.3.1 Network Model 603.3.2 Attack Model 603.3.3 Defense Model 623.4 CYBER-MIRA Framework 643.4.1 Architecture of CYBER-MIRA 653.4.2 Hypergame Expected Utility 743.4.3 Resilience Assessment as a Measure of Performance 823.5 Limitations 863.6 Conclusion and Future Work 863.6.1 Summary of the Key Contributions 863.6.2 Future Work 87References 874 Modeling Autonomous Network Resilience in Adversarial Environments Using Machine Learning and Topological Data Analysis 91Nandi O. Leslie4.1 Introduction 914.2 TDA Concepts 934.2.1 Simplices, Simplicial Complexes, and Filtration 934.2.2 Persistent Homology 954.3 Network Resilience Modeling 974.4 Conclusion 101References 1015 Game-theoretic Frameworks for Zero-trust Authentication in Autonomous Cyber Resilience 105Yunfei Ge and Quanyan Zhu5.1 Introduction 1055.2 From Traditional Security to Zero Trust 1085.3 Trust Evaluation Design 1105.3.1 Target 1105.3.2 Metric 1105.3.3 Collection and Evaluation 1115.3.4 Purpose 1155.3.5 Management 1155.4 Policy Engine Design 1165.4.1 Authentication Layer: Continuous Authentication 1175.4.2 Authorization Layer: Least Privilege Access 1175.4.3 Network Layer: Microsegmentation 1175.5 Zero Trust for Cyber Resilience 1185.5.1 How Zero Trust Contributes to Cyber Resilience 1185.5.2 A Running Example 1205.6 Strategic Zero-trust Implementation 1235.6.1 A Game-theoretic Approach 1235.6.2 Case Studies 1245.7 Conclusion 132References 1346 Cyber Insurance for Cyber Resilience 139Shutian Liu and Quanyan Zhu6.1 Introduction 1396.2 Attack Models and Insured Targets 1446.2.1 Human-layer Attacks 1446.2.2 Cyber-layer Attacks 1466.2.3 Physical-layer Attacks 1486.3 Defense Mechanisms and Residual Risks 1496.3.1 Modeling of Defense Mechanisms 1496.3.2 Types of Security Investments 1506.3.3 Residual Risk and Its Connection with Cyber Insurance 1536.4 Insurer’s Observations and the P-A Model 1556.4.1 User Behavior Monitoring 1556.4.2 Principal-agent Problems 1566.5 Modeling of Risk Preferences 1586.5.1 Risk Modeling 1596.5.2 Enhancing Cyber Resilience 1616.6 Insurance Design with Preference Manipulation 1626.7 Dynamic Insurances 1646.8 Regulations on Cyber Insurance 1666.8.1 Mandatory Cyber Insurance 1666.8.2 Insurance Policy Elaboration 1676.8.3 Designing Accountability Mechanisms 1686.8.4 Insurance Market Monitoring 1686.9 Conclusion 170References 1707 Enhancing Cyber Resiliency: Assessing the Effectiveness of Deploying Honeypots in Different Network Topologies 1837.1 Introduction 1837.1.1 Contribution 1847.2 Task Description 1857.2.1 Experiment Conditions 1857.2.2 Experiment Scenario 1877.2.3 Results 1897.2.4 Scanning and Exploitation Behavior 1897.2.5 Operating System and Exploit Preference 1927.3 A Cognitive Model of Attackers in HackIT 1937.3.1 IBL Theory 1937.3.2 IBL Model for Attacker 1957.3.3 IBL Model Results 1977.4 Discussion 199References 200Part 2 Resilient Machine Learning 2038 Computational Game Theory for Security 205Yevgeniy Vorobeychik8.1 Introduction 2058.2 Stackelberg Games 2068.3 Stackelberg Security Games 2098.4 Security Games on Networks 2128.5 Stochastic Stackelberg Games and Adversarial Patrolling 2158.5.1 Stochastic Discounted Stackelberg Games 2158.5.2 Adversarial Patrolling Games 2178.5.3 Solving Zero-sum APGs 2188.5.4 Solving General-sum APGs 2208.6 Conclusion 221References 2219 Privacy and Robustness Trade-offs of Artificial Intelligence Models with Federated Learning 225Kemal Davaslioglu, Yi Shi, and Yalin E. Sagduyu9.1 Introduction 2259.2 Model Inversion Attacks 2299.2.1 Softmax Regression Model Training 2299.2.2 ModInv Attacks Against a Single-layer Convolutional Neural Network (CNN) 2319.2.3 ModInv Attacks Against a Four-layer CNN 2319.2.4 Comparison of the Three Types of Models Under Attack 2329.2.5 Privacy Preservation Evaluation of Image Transformations Against ModInv Attacks 2339.2.6 Accuracy of These Privacy-preserving Image Transformations in ModInv Attacks 2419.2.7 Demonstration of ModInv Attacks Against CIFAR- 10Dataset 2429.3 Membership Inference Attacks 2439.3.1 Introduction 2439.3.2 General MI Attack Model 2469.3.3 Logistic Attack Models 2479.3.4 Overview of the MI Attack 2479.3.5 Naïve Attacks 2489.3.6 The Threat Models 2489.3.7 Naïve Bayes mi 2489.3.8 mi in Deep Models 2499.3.9 Differentially Private Stochastic Gradient Descent for Privacy 2529.3.10 Design Defense Approaches for MI Attack 2539.3.11 Membership Privacy in ml 2549.4 Federated Learning 2649.4.1 Federated Learning Implementation 2659.4.2 Aircraft Classification in the xView Dataset Using FL 2669.4.3 Effect of FL Parameters 2689.4.4 Demonstration of ModInv Attack Against FL Models 2709.4.5 Differentially Private Stochastic Gradient Descent 2719.4.6 Evaluate Effects of Different Parameters of DP-SGD 2729.4.7 Demonstration of Renyi DP Evaluations 2739.5 Discussion 2759.6 Conclusions 2769.7 Acknowledgment 277References 27710 Resilient Deep Neural Network Random Ensemble Against Adversarial Attacks 281Kirsen Sullivan, Yitao Li, Charles A. Kamhoua, and Bowei xi10.1 Introduction 28110.1.1 Related Work 28210.2 Data and Bootstrapped CNNs 28410.2.1 Bootstrap Three-layer CNN 28510.2.2 Bootstrap VGG 16 28710.2.3 Bootstrap Inception V 3 28810.3 Bootstrapped Distributions 29010.3.1 CNN3 Parameter 29010.3.2 VGG16 Parameters 29210.3.3 Inception v3 Parameters 29310.3.4 Normality Test 29610.3.5 NNs by Varying Initial Random Seeds 29710.3.6 Regression for NN Parameters 29910.4 Randomized DNN Ensembles with Gaussian Random Weights 30010.4.1 Adversarial Examples Generation 30010.4.2 Randomization 30110.5 Ensemble Experiment Results 30210.5.1 CNN3 Randomization Results 30210.5.2 VGG16 Randomization Results 30310.5.3 Inception v3 Randomization Results 30410.6 Conclusion 311References 312Part 3 Game Theory for Network Resilience 31711 Poisoning Attack and Defense Game for Federated Learning in Resilient NextG Networks 319Yalin E. Sagduyu, Tugba Erpek, and Yi Shi11.1 Introduction 31911.2 Federated Learning for Distributed Spectrum Monitoring 32311.3 Attack and Defense Mechanisms for Resilient Federated Learning 32511.4 Poisoning Attack–Defense Game for Two Clients 33011.5 Poisoning Attack–Defense Game for More than Two Clients 33711.6 Future Research Directions 33811.7 Conclusion 341References 34112 Self-adapting Quantum Network Provisioning Using Game Theory 347Stefan Rass, Miralem Mehic, Sandra König, Stefan Schauer, and Miroslav Voznak12.1 Introduction 34712.2 Basics of Quantum Networks 34812.3 Game Theory to Orchestrate Cryptography 34912.3.1 Basics of Perfectly Secure Message Transmission 35012.3.2 Hierarchical Secret Sharing and Access Structures 35212.3.3 Using Secret Sharing for Perfectly Secure Multipath Transmission (Defense Strategies) 35312.3.4 How to Define or Identify Adversary Structures (Attack Strategies)? 35412.3.5 Quantum Cryptography in Combination with Multipath Transmission 35412.3.6 Game-theoretic Orchestration of Perfectly Secure Message Transmission 35512.4 Self-adaption of QKD Devices to Environmental Conditions 36112.5 Adapting the Level of Service to Traffic Changes 36912.6 Adapting the Network Topology 37212.7 Conclusions, Outlook, and QKD in Today’s Networks 376References 37713 Conclusion and Future Works 383Quanyan Zhu13.1 Overview 38313.2 Summary 38413.2.1 Summary and Synthesis of Part 1 on Foundations of Cyber Resilience 38513.2.2 Summary and Synthesis of Part 2 on Resilient ml 39013.2.3 Summary of Part 3 on Game Theory for Network Resilience 39313.3 Future Directions: Charting the Path Toward Autonomous Cyber Resilience 39613.3.1 Learning Autonomy: From Reactive Defense to Strategic Adaptation 39613.3.2 Multiscale Resilience: Tailoring Autonomy Across Layers and Contexts 39713.3.3 Game-theoretic Intelligence: Strategic Reasoning for Resilience 39913.3.4 Holistic and Integrative Approaches: Building a Converged Resilience Architecture 40113.3.5 Research and Policy Translation: From Innovation to Implementation 40313.4 Concluding Remarks 404Index 407