Security Threats and Defense Systems Architectures for Multimodal Large Language Models
定价：￥159.00
作者： 谢盈，丁旭阳，张小松
出版时间：2026-06
出版社：科学出版社
以下为《Security Threats and Defense Systems Architectures for Multimodal Large Language Models》的配套数字资源，这些资源在您购买图书后将免费附送给您：
- 科学出版社
- 9787030848901
- 1版
- 2026-06
作者简介
内容简介
Against the backdrop of the technological development of multimodal models, this book first introduces its core architecture and typical application scenarios, building upon this foundation to construct a security threat analysis framework covering the entire life cycle. It then delves into typical attack methods faced during the training and inference phases, such as data poisoning, backdoor attacks, cross?鄄modal adversarial examples, and prompt injection, further proposing systematic defense strategies and engineering implementation methods covering all stages of model training, inference, and deployment. Finally, in conjunction with cutting?鄄edge technological developments, it provides an outlook on the future trends and challenges of multimodal model security.
目录
Contents
Part Ⅰ MLLMs: Technology and Security Overview 1
Chapter 1 Introduction to MLLMs: Technology and Security 3
1.1 Overview of MLLM Technology 3
1.1.1 From Unimodal LLMs to MLLMs: Capability and Architecture Shifts 4
1.1.2 Comparative Analysis of Mainstream Multimodal Model Structures 8
1.1.3 Typical Application Scenarios of Multimodal Models 16
1.2 Overview of the Life Cycle and Attack Surface of MLLMs 22
1.2.1 Overview of the Life Cycle of MLLMs 22
1.2.2 Distribution of Potential Security Risks at Each Stage of the Model 28
1.2.3 Analysis of New Risk Types Specific to Multimodality 39
1.3 Chapter Summary 47
References 47
Part Ⅱ Overview of Threat Models and Attack Techniques for Multimodal Models 49
Chapter 2 Overview of Threat Models and Attack Techniques for Multimodal Models 51
2.1 Data and Training Phase Attack Techniques 51
2.1.1 Modality-Aware Data Poisoning Attacks 52
2.1.2 Cross-Modal Backdoor Attack Mechanism 60
2.1.3 Model Inversion and Membership Inference Attacks 69
2.1.4 Modality Imbalance Attacks 75
2.2 Attack Techniques in the Reasoning and Fusion Phase 81
2.2.1 Image-Text Joint Adversarial Sample Attacks 82
2.2.2 Cross-Modal Hallucination Attacks 88
2.2.3 Modal Injection Attacks 96
2.3 Risks of Abuse and External Attacks 102
2.3.1 Risks of Abuse 103
2.3.2 External Attacks 112
2.4 Chapter Summary 121
References 122
Part Ⅲ Overview of Defense Strategies for MLLMs 125
Chapter 3 Introduction to MLLMs Defense Strategies 127
3.1 Multimodal Adversarial Defense Strategies 127
3.1.1 Multimodal Adversarial Training Strategy 128
3.1.2 Modal Weight Balancing Mechanism 136
3.1.3 Backdoor Detection and Modal Trigger Location Method 145
3.1.4 Adaptability of Differential Privacy Mechanisms in Multimodal Training 154
3.2 Defense Strategies for the Inference and Interface Phases 163
3.2.1 Prompt Injection Defense Mechanism 164
3.2.2 Purification and Cross-Modal Consistency Detection 171
3.2.3 Hallucination Detection in MLLMs 179
3.3 Security Strategy for Deployment and Interaction Stages 185
3.3.1 Model Security Review and Deployment 186
3.3.2 Red-Blue Adversarial Security Testing System 192
3.3.3 Model Watermarking and Generation Traceability Mechanism 200
3.4 Chapter Summary 205
References 206
Part Ⅳ Overview of Security Challenges and Future Development Trends 209
Chapter 4 Security Challenges and Future Development Trends 211
4.1 Outlook on Technology Development Trends 211
4.1.1 Adaptive Modality Selection and Input Optimization Mechanism 212
4.1.2 Security Enhancement System for Multimodal Models 215
4.1.3 Multimodal Security Protection on Edge Devices 219
4.2 Security Alignment and Social Governance 226
4.2.1 Challenges of Multimodal Secure Alignment: RLHF Applications in the Visual
Modality 227
4.2.2 Construction and Implementation Difficulties of MLLMs Security Governance
Framework 232
4.3 Chapter Summary 235
References 235
Part Ⅰ MLLMs: Technology and Security Overview 1
Chapter 1 Introduction to MLLMs: Technology and Security 3
1.1 Overview of MLLM Technology 3
1.1.1 From Unimodal LLMs to MLLMs: Capability and Architecture Shifts 4
1.1.2 Comparative Analysis of Mainstream Multimodal Model Structures 8
1.1.3 Typical Application Scenarios of Multimodal Models 16
1.2 Overview of the Life Cycle and Attack Surface of MLLMs 22
1.2.1 Overview of the Life Cycle of MLLMs 22
1.2.2 Distribution of Potential Security Risks at Each Stage of the Model 28
1.2.3 Analysis of New Risk Types Specific to Multimodality 39
1.3 Chapter Summary 47
References 47
Part Ⅱ Overview of Threat Models and Attack Techniques for Multimodal Models 49
Chapter 2 Overview of Threat Models and Attack Techniques for Multimodal Models 51
2.1 Data and Training Phase Attack Techniques 51
2.1.1 Modality-Aware Data Poisoning Attacks 52
2.1.2 Cross-Modal Backdoor Attack Mechanism 60
2.1.3 Model Inversion and Membership Inference Attacks 69
2.1.4 Modality Imbalance Attacks 75
2.2 Attack Techniques in the Reasoning and Fusion Phase 81
2.2.1 Image-Text Joint Adversarial Sample Attacks 82
2.2.2 Cross-Modal Hallucination Attacks 88
2.2.3 Modal Injection Attacks 96
2.3 Risks of Abuse and External Attacks 102
2.3.1 Risks of Abuse 103
2.3.2 External Attacks 112
2.4 Chapter Summary 121
References 122
Part Ⅲ Overview of Defense Strategies for MLLMs 125
Chapter 3 Introduction to MLLMs Defense Strategies 127
3.1 Multimodal Adversarial Defense Strategies 127
3.1.1 Multimodal Adversarial Training Strategy 128
3.1.2 Modal Weight Balancing Mechanism 136
3.1.3 Backdoor Detection and Modal Trigger Location Method 145
3.1.4 Adaptability of Differential Privacy Mechanisms in Multimodal Training 154
3.2 Defense Strategies for the Inference and Interface Phases 163
3.2.1 Prompt Injection Defense Mechanism 164
3.2.2 Purification and Cross-Modal Consistency Detection 171
3.2.3 Hallucination Detection in MLLMs 179
3.3 Security Strategy for Deployment and Interaction Stages 185
3.3.1 Model Security Review and Deployment 186
3.3.2 Red-Blue Adversarial Security Testing System 192
3.3.3 Model Watermarking and Generation Traceability Mechanism 200
3.4 Chapter Summary 205
References 206
Part Ⅳ Overview of Security Challenges and Future Development Trends 209
Chapter 4 Security Challenges and Future Development Trends 211
4.1 Outlook on Technology Development Trends 211
4.1.1 Adaptive Modality Selection and Input Optimization Mechanism 212
4.1.2 Security Enhancement System for Multimodal Models 215
4.1.3 Multimodal Security Protection on Edge Devices 219
4.2 Security Alignment and Social Governance 226
4.2.1 Challenges of Multimodal Secure Alignment: RLHF Applications in the Visual
Modality 227
4.2.2 Construction and Implementation Difficulties of MLLMs Security Governance
Framework 232
4.3 Chapter Summary 235
References 235
