Privacy and Security for Large Language Models. Hands-On Privacy-Preserving Techniques for Personalized AI Baihan Lin

Dostępność produktu

Produkt nie został jeszcze oceniony pod kątem ułatwień dostępu lub nie podano żadnych informacji o ułatwieniach dostępu lub są one niewystarczające. Prawdopodobnie Wydawca/Dostawca jeszcze nie umożliwił dokonania walidacji produktu lub nie przekazał odpowiednich informacji na temat jego dostępności.

Spis treści książki

• Preface
  • Who Should Read This Book
  • Why I Wrote This Book
  • Navigating This Book
  • Conventions Used in This Book
  • Using Code Examples
  • OReilly Online Learning
  • How to Contact Us
  • Acknowledgments
• 1. Introduction
  • The Rise of Large Language Models
  • Privacy and Security Concerns in LLMs
  • What This Book Covers
  • Your Role in This Journey
  • Summary
• 2. Understanding Large Language Models
  • Fundamentals of Large Language Models
    • Basic Building Blocks of Language Models
      • Neural networks
      • Recurrent neural networks
      • Long short-term memory networks
    • Key Concepts in LLMs
      • Tokenization
      • Chunking
      • Embeddings
      • Attention mechanisms
      • Context windows and sequence length
      • Transfer learning and foundation models
      • Discriminative versus generative models
      • In-context learning
      • Zero-shot and few-shot learning
  • LLM Architectures
    • Transformer Architecture
    • Mixture of Experts Architecture
    • Popular LLM Models
  • Training Techniques for LLMs
    • Pre-Training Techniques
      • Masked language modeling
      • Next sentence prediction
      • Permutation Language Modeling
    • Fine-Tuning Techniques
      • Full fine-tuning
      • Parameter-efficient fine-tuning
      • Instruction fine-tuning
      • Reinforcement Learning from Human Feedback
  • Retrieval-Augmented Generation
  • Summary
• 3. Evaluating the Privacy and Security Risks of LLMs
  • Privacy Metrics
    • Differential Privacy
      • Mathematical formulation
      • Code implementation
    • Privacy Loss
      • Mathematical formulation
      • Code implementation
    • k-anonymity
      • Code implementation
    • Privacy Considerations in RAG Systems
  • Security Metrics
    • Attack Success Rate (ASR)
      • Code implementation
    • False Positive Rate (FPR) for Membership Inference
      • Code implementation
    • Reconstruction Error for Model Inversion
      • Code implementation
  • LLM Privacy and Security Audits
    • Simulating Attacks
      • Membership inference attack
        • Operational definition 1: Perplexity-based attack
        • Operational definition 2: Repeated prompt-based attack
        • What does this mean?
      • Data extraction attack
        • Operational definition: Prompt-based extraction
        • How does this attack work?
        • Code example: Generate text for data extraction
        • Enhancing the attack
        • Defense: How to guard against data extraction
    • LLMPrivacySecurityEvaluator: The All-in-One Auditor
      • Expanding the LLMPrivacySecurityEvaluator
      • Interpreting the results
      • Different personas, different approaches
  • Modern Evaluation Frameworks and Benchmarks
  • Summary
• 4. Privacy-Preserving Training Techniques
  • A Real-World Example of Privacy Breach in the Training Phase
    • Synthetic Data for Privacy Evaluation
      • Why synthetic data?
    • How to Apply LLMPrivacySecurityEvaluator on Your Data
  • Differential Privacy for LLMs
    • The Mathematical Foundation
    • Implementing DP-SGD for LLMs
    • Privacy Accounting in Practice
    • Trade-Offs and Considerations
    • Applying Differential Privacy to Retrieval-Augmented Generation
  • Federated Learning with LLMs
    • The Concept
    • Implementing Federated Learning for LLMs
    • Advantages and Challenges of Federated Learning
  • Homomorphic Encryption in LLMs
    • The Concept
    • Implementing HE for LLMs
    • Advantages and Challenges of Homomorphic Encryption
  • Multi-Party Computation for Secure Aggregation
    • The Concept
    • Implementing MPC with Modern Libraries
    • Advantages and Challenges of MPC
  • Parameter-Efficient Fine-Tuning for Privacy
    • Low-Rank Adaptation
    • Quantized Low-Rank Adaptation
  • Privacy-Preserving Data Transformation
    • Data Anonymization and De-Identification
    • Privacy-Preserving Data Augmentation
    • Advantages and Challenges of Privacy-Preserving Data Augmentation
  • Summary
• 5. Secure Deployment of LLMs
  • Secure Model Hosting and Infrastructure
    • Understanding Infrastructure Components
    • Isolation Strategies
      • Containerization
      • Virtual machines
    • Network Security
      • Network architecture design
      • HTTPS and TLS implementation
    • Resource Management and Monitoring
  • Secure APIs and Communications
    • API Design Principles
    • Implementation of Secure APIs
    • Authentication and Authorization
    • Secure Communication
      • Symmetric encryption: AES-256
      • Network-level encryption: WPA3
      • Implementing encryption standards in LLMs
  • Secure Model Versioning and Updates
    • Model Registry and Version Control
    • Secure Update Process
  • Summary
• 6. Adversarial Attacks and Defenses
  • Understanding Adversarial Attacks on LLMs
    • Taxonomy of Adversarial Attacks on LLMs
    • Notable Attack Methods
      • Jailbreaking attacks
      • Adversarial prompts generation
      • Universal adversarial triggers
    • Embedding Space Attacks
    • LLM Agent Attacks
    • Impact of Model Scale and Architecture
    • Case Study: Defending Against Jailbreaking Attacks
  • Robust Fine-Tuning Techniques
    • Adversarial Training
    • Robust Optimization Techniques
      • Misclassification Aware Regularization Technique (MART)
      • TRade-offinspired Adversarial DEfense via Surrogate-loss minimization (TRADES)
    • Data Augmentation for Robustness
    • Prefix-Tuning and Prompt-Based Robustness
    • Ensemble Methods
    • Certifiably Robust Fine-Tuning
  • Red-Teaming LLMs
    • Red-Teaming Methodologies
      • Manual red-teaming
      • Automated red-teaming
    • Implementing a Red-Teaming Program
    • Red-Teaming Tools and Frameworks
    • Automated Multiround Red-Teaming
    • Case Study: Red-Teaming in Practice
  • Adversarial Evaluation and Robustness Metrics
    • Robustness Benchmarks
    • Robustness Under Distribution Shift
    • Human-in-the-Loop Evaluation
    • Agent-Based Evaluation
    • Standardized Attack Success Metrics
    • Defense Evaluation Metrics
    • Challenges in Robustness Evaluation
  • Best Practices
  • Future Directions in LLM Robustness
  • Summary
• 7. Ethical Considerations in Fine-Tuning LLMs
  • Bias and Fairness Issues in Personalization
    • Understanding Bias in Fine-Tuned LLMs
    • Measuring Fairness in Fine-Tuned Models
      • Demographic parity (statistical parity)
      • Equalized odds (error rate balance)
      • Individual fairness (Lipschitz fairness)
      • Calibration fairness
      • Counterfactual fairness
    • Bias Mitigation Strategies
    • Challenges in Privacy-Preserving Bias Mitigation
  • Transparency and Explainability in Fine-Tuned Models
    • The Explainability Challenge in LLMs
    • Techniques for Explaining LLM Behavior
    • Privacy-Preserving Explainability
  • Addressing AI Bias with Privacy Constraints
    • The Privacy-Fairness Trade-Off
    • Group-Aware Privacy Mechanisms
    • Bias-Aware Federated Learning
    • Privacy-Preserving Bias Auditing
  • Summary
• 8. Navigating the Cultural, Social, and Legal Landscapes
  • A New Kind of Socio-Technical Systems
  • Riding Amidst an AI-Mediated Cultural Evolution
    • The Rise of AI-Generated Content and the Erosion of Trust
    • Personalized AI and Identity Crisis in the Age of Surveillance Capitalism
    • Existential Questions in Human-Machine Interaction
    • Unveiling the Generative AI Supply Chain
    • The Emergence of Machine Culture
  • Adaptable Legal Frameworks for Regulation and Accountability
    • The Case of Copyright and Intellectual Property in the Age of LLMs
    • The Case of Data Privacy and Protection in Personalized AI Systems
    • The Case of Algorithmic Bias and Discrimination in AI-Powered Decision Making
    • The Case of Liability and Accountability in AI-Powered Systems
    • Universal Challenges to Techno-Legal Solutionism
  • Building a Responsible AI Culture
  • AI Safety Beyond Algorithms: The Human Elements
  • Summary
• 9. Building Privacy-Preserving AI Capabilities
  • Healthcare AI in Action: Differentially Private Clinical Note Analysis
    • The Healthcare Privacy Challenge
    • Synthetic Data as a Privacy-Preserving Foundation
    • LoRA: Efficient and Privacy-Friendly Fine-Tuning
    • Privacy Accounting with RDP
    • Real-World Deployment Considerations
  • Legal AI in Action: Federated Learning Across Law Firms or Courts
    • The Legal Confidentiality Imperative
    • Federated Learning Architecture for Legal AI
    • Secure Aggregation and Model Updates
    • Legal and Ethical Considerations in Federated Legal AI
    • Performance and Utility Evaluation
  • Building Your Privacy-First AI Capability
    • Organizational Readiness and Implementation Strategy
    • Team Structure and Technology Decisions
    • Governance Integration and Success Measurement
  • Preparing for Tomorrows Privacy Landscape
    • Technology Convergence and Regulatory Evolution
    • Market Dynamics and Competitive Positioning
    • A Strategic Position for the Future
  • Summary
  • Conclusion
    • The Transformation Youve Witnessed
    • The Path Were On
    • Your Role in Shaping the Future
• Index

pokaż cały spis treści

ukryj recenzje