Privacy in ML: anonymisation, differential privacy, DP-SGD

11 steps · about 8 h 9 min · Intermediate

  1. Privacy-Preserving Machine Learning: A Practical Guide (guide, Alexandra Institute, 30 min)

    It maps the whole field first: the privacy risks, anonymisation, membership inference and the main defences.

  2. Privacy Preserving Machine Learning, Lecture 1: Introduction & Course Overview (doc, Aurélien Bellet, Inria, 40 min)

    It shows why stripping identifiers and k-anonymity break down, which motivates differential privacy.

  3. Privacy Preserving Machine Learning (PPML) @ SciPy 2023 (doc, Valerio Maggio, Speaker Deck, 25 min)

    It compares k-anonymity with differential privacy and lists the attacks on ML models, such as model inversion and membership inference.

  4. Introduction to Differential Privacy in Deep Learning Models (article, Baeldung, 15 min)

    A short plain-language explanation of the core idea: one person's record should barely change the output.

  5. Tutorial #12: Differential Privacy I: Introduction (blog, RBC Borealis, 40 min)

    It builds the formal definition, randomized response and Laplace noise step by step.

  6. A Short Tutorial on Differential Privacy (video, Borja Balle, Alan Turing Institute (via Class Central), 1 h 49 min)

    A full lecture that consolidates the definitions, mechanisms and how they combine.

  7. Differential Privacy Overview and Fundamental Techniques (reference, arXiv, 1 h)

    A rigorous chapter on why earlier privacy attempts failed and on the standard DP mechanisms, aimed at AI use.

  8. Machine Learning with Differential Privacy in TensorFlow (tutorial, TensorFlow Privacy, 45 min)

    It walks through DP-SGD (clipping gradients, adding noise) on MNIST with the original library.

  9. Understanding Differential Privacy: Part 2 (blog, Hexens, 40 min)

    It trains a baseline and a DP model with Opacus side by side, so you see the accuracy cost and the epsilon budget.

  10. Opacus Tutorials (interactive, Opacus (Meta), 1 h)

    It offers the official PyTorch DP-SGD blog series and notebooks to try on your own models.

  11. Differential Privacy for AI: Protecting Training Data (2026) (guide, AI Security and Safety, 25 min)

    It ends with a practical checklist: tool choice, LLM fine-tuning with DP, and regulation.