Hands-On Differential Privacy Introduction to the Theory and Practice Using OpenDP (Ethan Cowan, Michael Shoemate etc.)（Z-Library）

# Hands-On Differential Privacy: Introduction to the Theory and Practice Using OpenDP ## 【One-Line Pitch】 A practical, code-first introduction to differential privacy that takes you from the core mathematical concepts to real deployments using the OpenDP library—ideal for data scientists, researchers, and engineers who handle sensitive datasets and need more than anonymization can offer. ## 【Book Arc】 - **Opening (~0%–9%)**: Sets the stage with the failure of traditional deidentification and anonymization methods, introduces the history of data privatization, and walks through a classroom exam-score example that builds intuition for key terms like sensitivity and adjacent datasets. - **Early (~9%–25%)**: Formalizes differential privacy—defining privacy loss, the formal DP guarantee, randomized response, and the central vs. local models of privacy—while introducing the Laplace mechanism and accuracy guarantees. - **Early–Middle (~25%–38%)**: Develops the foundational theory of stable transformations, c-stability, and metric spaces, showing how sensitivity emerges as a special case of stability and proving stability for basic functions like count, sum, and double. - **Middle (~38%–47%)**: Extends stability analysis to vector-valued aggregators, Lp norms, and bounded-norm settings, deriving sensitivities for sum, mean, and variance under different assumptions about dataset size and bounds. - **Middle (~47%–end of sample)**: Bridges from transformations to private mechanisms, replacing output metrics with privacy measures and generalizing the definition of privacy—the conceptual pivot that connects all prior theory to actual DP implementations. ## 【Key Takeaways】 - **Traditional anonymization is insufficient** (Opening): Deidentification and related statistical disclosure limitation techniques fail against modern attacks; DP provides formal, provable guarantees that anonymization cannot. This motivates the entire book. - **Sensitivity quantifies the worst-case impact of one individual** (Early): The global sensitivity of a statistic—like the maximum change in a classroom mean when one score is removed—determines how much noise is needed. The book emphasizes proving sensitivity mathematically rather than relying on ad hoc reasoning. - **Central vs. local models involve a trust/utility trade-off** (Early): The central model privatizes data after collection on a trusted authority, yielding better utility; the local model has individuals privatize their own data, requiring significantly more data for the same guarantee. Rule of thumb: privatize as late as possible. - **The Laplace mechanism provides calibrated noise with accuracy guarantees** (Early): For a given sensitivity, Laplace noise scaled appropriately yields (α, β)-accuracy, enabling confidence intervals around private outputs—demonstrated with a student family-size counting example. - **Stability is the unifying concept for DP** (Early–Middle): A transformation is c-stable if neighboring inputs map to neighboring outputs within a bounded distance; sensitivity is a special case of stability. This abstraction lets you reason about privacy systematically. - **Unbounded functions require input constraints** (Middle): The sum function is not stable without bounds on data values; adding assumptions like L and U bounds (or radius R and origin O) makes stability derivable and enables privacy guarantees. - **Private mechanisms generalize stable transformations** (Middle): The same structure applies, but the output metric is replaced by a privacy measure—making the definition of privacy itself a specialized kind of stability constraint. ## 【Reading Tips】 - **Read Chapters 1–2 carefully if you're new to DP** (per the authors' own recommendation): These establish the vocabulary—adjacent datasets, sensitivity, privacy loss, local vs. central models—that everything else builds on. Don't skip the classroom example; it grounds abstract concepts. - **Work through the stability proofs with pen and paper** (Early–Middle): The derivations for count, sum, and double transformations follow a repeatable pattern (start with the left side, simplify, identify the constant c). Replicating them yourself builds the mathematical maturity needed for later chapters. - **Skim the exercises for self-assessment** (throughout): Each chapter ends with conceptual, theoretical, and programming questions. The adjacency-detection and stability-derivation exercises are particularly good checks of understanding before moving forward. - **Expect a shift from intuition to formalism around the 25–40% mark**: The book deliberately becomes more abstract with metric spaces and c-stable transformations. If this feels heavy, remember the payoff: these tools let you derive privacy guarantees for any new function you encounter. - **Use the OpenDP code examples as anchors** (throughout): When theory gets dense, the library demonstrations—like computing stability for bounded datasets—show the practical application and help cement what the math means in practice. ## 【Coverage Limits】 This guide covers the conceptual and theoretical foundations visible in the sampled excerpts (roughly the first half of the book). The later parts—synthetic data generation, privacy attacks, setting privacy loss parameters, and project planning—are mentioned in the table of contents but not covered in detail here. ##