Hands-On Differential Privacy

Spis treści książki

• Preface
  • The Structure of This Book
    • Part 1: Differential Privacy Concepts
    • Part 2: Differential Privacy in Practice
    • Part 3: Deploying Differential Privacy
  • Conventions Used in This Book
  • Using Code Examples
  • OReilly Online Learning
  • How to Contact Us
  • Acknowledgments
• I. Differential Privacy Concepts
• 1. Welcome to Differential Privacy
  • History
  • Privatization Before Differential Privacy
  • Case Study: Applying DP in a Classroom
    • Privacy and the Mean
    • How Could This Be Prevented?
      • Randomized response
      • Adding noise
  • Adjacent Data Sets: What If Someone Else Had Dropped the Class?
  • Sensitivity: How Much Can the Statistic Change?
  • Adding Noise
    • What Is a Trusted Curator?
  • Available Tools
  • Summary
  • Exercises
• 2. Differential Privacy Fundamentals
  • Intuitive Privacy
    • Privacy Unit
    • Privacy Loss
  • Formalizing the Concept of Differential Privacy
    • Randomized Response
      • Randomized response is -DP
    • Privacy Violation
  • Models of Differential Privacy
  • Sensitivity
  • Differentially Private Mechanisms
    • Laplace Mechanism
    • The Laplace Mechanism Is -DP
    • Mechanism Accuracy
    • Most Common Family Type Among Students
    • Exponential Mechanism
  • Composition
  • Postprocessing Immunity
  • Implementing Differentially Private Queries with SmartNoise
    • Example 1: Differentially Private Counts
    • Example 2: Differentially Private Sum
    • Example 3: Multiple Queries from a Single Database
  • Summary
  • Exercises
• 3. Stable Transformations
  • Distance Metrics
    • Data Set Adjacency
    • Bounded Versus Unbounded Differential Privacy
  • Definition of a c-Stable Transformation
    • Transformation: Double
    • Transformation: Row-by-Row
  • Stability Is a Necessary and Sufficient Condition for Sensitivity
    • Transformation: Count
    • Transformation: Unknown-Size Sum
  • Domain Descriptors
    • Transformation: Data Clipping
  • Chaining
  • Metric Spaces
  • Definition of Stability
    • Transformation: Known-Size Sum
    • Transformation: Known-Size Mean
    • Transformation: Unknown-Size Mean
    • Transformation: Resize
    • Recap of Scalar Aggregators
  • Vector-Valued Aggregators
    • Vector Norm, Distance, and Sensitivity
    • Aggregating Data with Bounded Norm
    • Grouped Data
  • In Practice
  • Summary
  • Exercises
• 4. Private Mechanisms
  • Privacy Measure
    • Privacy Measure: Max-Divergence
    • Metric Versus Divergence Versus Privacy Measure
  • Private Mechanisms
    • Randomized Response
    • The Vector Laplace Mechanism
      • Discrete Laplace mechanism (geometric mechanism)
    • Exponential Mechanism
    • Quantile Score Transformation
      • Finite support exponential mechanism
      • Piecewise-Constant Support Exponential Mechanism
      • Top-K
    • Report Noisy Max Mechanisms
      • Alternative distributions
  • Interactivity
  • Above Threshold
    • Streams
    • Online Private Selection
    • Stable Transformations on Streams
  • Summary
  • Exercises
• 5. Definitions of Privacy
  • The Privacy Loss Random Variable
  • Approximate Differential Privacy
    • Truncated Noise Mechanisms
    • Propose-Test-Release
    • (Advanced) Composition
  • The Gaussian Mechanism
  • Rényi Differential Privacy
    • Zero-Concentrated Differential Privacy (zCDP)
    • Strength of Moments-Based Privacy Measures
  • Bounded Range
  • Privacy Loss Distributions
    • Numerical Composition
    • Characteristic Functions
  • Hypothesis Testing Interpretation
    • f -differential privacy
  • Summary
  • Exercises
• 6. Fearless Combinators
  • Chaining
    • Example: Bounds Estimation
    • Example: B-Tree
  • Privacy Measure Conversion
  • Composition
    • Adaptivity
    • Odometers and Filters
  • Partitioned Data
    • Example: Grouping on Asylum Seeker Data
    • Parallel Composition
    • Example: Multi-Quantiles
  • Privacy Amplification
    • Privacy Amplification by Simple Random Sampling
    • Privacy Amplification by Poisson Sampling
    • Privacy Amplification by Shuffling
  • Sample and Aggregate
  • Private Selection from Private Candidates
    • Example: k-Means
  • Summary
  • Exercises
• II. Differential Privacy in Practice
• 7. Eyes on the Privacy Unit
  • Levels of Privacy
    • User-Level Privacy in Practice
  • Browser Logs Example: A Naive Event-Level Guarantee
  • Data Sets with Unbounded Contributions
    • Statistics with Constant Sensitivity
  • Data Set Truncation
    • Reservoir Sampling
    • Truncation on Partitioned Data
    • Hospital Visits Example: A Bias-Variance Trade-Off
      • Scenario 1: Truncation threshold matches the true maximum influence of any single individual in the data
      • Scenario 2: Maximum allowed number of visits per patient overestimates the true maximum influence of any single individual in the data
      • Scenario 3: Maximum allowed number of visits per patient underestimates the true maximum influence of any single individual in the data
  • Privately Estimating the Truncation Threshold
    • Further Analysis with Unbounded Contributions
  • Unknown Domain
  • When to Apply Truncation
    • Stable Grouping Transformations
    • Stable Union Transformations
    • Stable Join Transformations
  • Summary
  • Exercises
• 8. Differentially Private Statistical Modeling
  • Private Inference
  • Differentially Private Linear Regression
    • Sufficient Statistics Perturbation
    • Private Theil-Sen Estimator
    • Objective Function Perturbation
  • Algorithm Selection
  • Differentially Private Naive Bayes
    • Categorical Naive Bayes
    • Continuous Naive Bayes
    • Mechanism Design
    • Example: Naive Bayes
  • Tree-Based Algorithms
  • Summary
  • Exercises
• 9. Differentially Private Machine Learning
  • Why Make Machine Learning Models Differentially Private?
  • Machine Learning Terminology Recap
  • Differentially Private Gradient Descent (DP-GD)
    • Example: Minimum Viable DP-GD
  • Stochastic Batching (DP-SGD)
    • Parallel Composition
    • Privacy Amplification by Subsampling
    • Hyperparameter Tuning
      • Public holdout
      • Private selection from private candidates
  • Private Aggregations of Teacher Ensembles
  • Training Differentially Private Models with PyTorch
    • Example: Predicting Income Privately
  • Summary
  • Exercises
• 10. Differentially Private Synthetic Data
  • Defining Synthetic Data
    • Types of Synthetic Data
  • Practical Scenarios for Synthetic Data Usage
  • Marginal-Based Synthesizers
    • Multiplicative Weights Update Rule with the Exponential Mechanism
  • Graphical Models
    • PrivBayes
  • GAN Synthesizers
    • Potential Problems
  • Summary
  • Exercises
• III. Deploying Differential Privacy
• 11. Protecting Your Data Against Privacy Attacks
  • Definition of a Privacy Violation
  • Attacks on Tabular Data Sets
    • Record Linkage
      • Example: teacher survey
    • Singling Out
    • Differencing Attack
    • Reconstruction via Systems of Equations
    • Tracing
    • k-Anonymity Vulnerabilities
  • Attacks on Machine Learning
  • Summary
  • Exercises
• 12. Defining Privacy Loss Parameters of a Data Release
  • Sampling
  • Metadata Parameters
  • Allocating Privacy Loss Budget
  • Practices That Aid Decision-Making
    • Codebook and Data Annotation
    • Translating Contextual Norms into Parameters
  • Making These Decisions in the Context of Exploratory Data Analysis
  • Adaptively Choosing Privacy Parameters
  • Potential (Unexpected) Consequences of Transparent Parameter Selection
  • Summary
  • Exercises
• 13. Planning Your First DP Project
  • DP Deployment Considerations
    • Frequency of DP Deployments
    • Composition and Budget Accountability
  • DP Deployment Checklist
  • An Example Project: Back to the Classroom
  • Proper Real-World Data Publications
    • LinkedIns Economic Graph
    • Microsofts Broadband Data
  • DP Release Table: A Standard for Releasing Details About Your Release
  • Thats All, Folks
• Further Reading
  • Theory
  • Applications
• A. Supplementary Definitions
• B. Rényi Differential Privacy
  • Theorem: RDP Is Immune to Postprocessing
    • Proof
  • Theorem: Youngs Inequality
    • Proof via Calculus
  • Elementary Proof
  • Theorem: Holders Inequality
    • Proof
  • Theorem: Probability Preservation
    • Proof
  • Theorem: RDP to ( , ) -DP
    • Proof
• C. The Exponential Mechanism Satisfies Bounded Range
  • Proof
• D. Structured Query Language (SQL)
• E. Composition Proofs
  • Theorem: Basic Sequential Composition
    • Proof
  • Theorem: General Sequential Composition
    • Proof
  • Theorem: Parallel Composition
    • Proof
  • Theorem: Immunity to Postprocessing
    • Proof
• F. Machine Learning
  • Supervised Versus Unsupervised Learning
  • Gradient Descent
    • Using Gradient Descent to Learn Parameters
    • Stochastic Gradient Descent
• G. Where to Find Solutions
• Index

pokaż cały spis treści

ukryj recenzje