A model trained on sensitive data can leak it: attackers can infer whether a record was in the training set, reconstruct attributes, or extract memorized text from a generative model. NIST AI 100-2 E2025 classes these as privacy attacks alongside evasion, poisoning and misuse. This course teaches the techniques that reduce that leakage and, just as important, how to tell what guarantee each one really provides.
You will learn differential privacy from first principles (epsilon, delta, composition), train and fine-tune with DP-SGD using Opacus, choose and evaluate privacy budgets with NIST SP 800-226 (final, March 2025), and generate differentially private synthetic data. You will examine federated learning with frameworks such as Flower, what secure aggregation does and does not protect, and attacks on federated updates. The cryptographic and hardware module compares homomorphic encryption, secure multi-party computation and confidential computing on CPUs and GPUs by cost, latency and trust assumptions. For LLM systems you will design de-identification before training and retrieval, privacy-preserving inference gateways, and assess machine unlearning. The course closes with LINDDUN privacy threat modeling, mapping techniques to GDPR and EU AI Act data governance duties, and auditing models with attack simulations. The course is educational and is not legal advice. Operational data security for RAG stores and logs is covered in sec-230. The Wirewalk Institute certificate attests to completion and assessment of this course; it is not endorsed by OWASP, MITRE, NIST or any vendor named here.