"Membership Inference & Privacy Leakage: Measuring and Reducing Data Exposure" Modern ML systems leak more than accuracy reports reveal—and membership inference has become the most practical lens for turning “privacy risk” into something you can test, debug, and ship against. This book targets experienced ML engineers, security engineers, and applied researchers who own real deployments: models behind APIs, embedding services, analytics pipelines, and iterative retraining loops where small interface choices can quietly expose sensitive training data. You’ll learn to threat-model privacy leakage with attacker-centric precision; connect generalization, stability, and data pathologies to concrete membership signals; and navigate the full attack landscape from score-based black-box probes to shadow-model transfer and white-box gradient attacks. The book then teaches how to measure leakage correctly—using operational metrics like TPR at low FPR, contamination-free dataset construction, reproducibility protocols, and ablations that prevent false security claims. From there, you’ll implement mitigations that work in practice (output and query controls, stability-oriented training, calibration trade-offs) and progress to differential privacy fundamentals, privacy accounting, and the engineering realities of DP-SGD. Prerequisites include strong ML foundations and comfort with experimentation rigor. The differentiator is an end-to-end, deployment-first approach: privacy acceptance criteria, regression gates, monitoring and incident response, and framework-agnostic tooling patterns