01Aug 24-28
Course overview and privacy in practice
ML background for privacy
How can we use data to build useful systems without exposing the people behind the data? This course introduces the core ideas of modern data privacy through concrete attacks, practical defenses, and hands-on system audits.
The course is designed for advanced undergraduates. We will start with privacy failures that students can observe directly, then build toward differential privacy, privacy-aware machine learning, and privacy-enhancing technologies such as MPC, HE, TEE, and network privacy tools. The emphasis is on technical understanding, experimental reasoning, and clear communication rather than graduate-level novelty.
This version of the course is aimed at advanced undergraduates in computer science, data science, or related areas.
Required background:
Recommended background:
You do not need prior experience with LLM training, privacy research, or advanced cryptography.
Privacy is now part of the job in machine learning, data science, and systems work. Engineers are expected to understand not only how to build models, but also how those models leak, what protections are realistic, and where the trade-offs appear in practice. This course is intended to prepare students for that level of technical judgment.
| Course documents: Syllabus | Policy |
Tentative Tuesday/Thursday plan following the UVA academic calendar. Reading Days, Election Day, Thanksgiving recess, and the December 8 course end date are reflected below.
Course overview and privacy in practice
ML background for privacy
Extraction and memorization
Membership inference and attack evaluation
Attack evaluation, baselines, and reproducibility
Linkage, singling-out, and reconstruction
Anonymization and its limits
DP definition, adjacency, and sensitivity
Async attack case study and Quiz 1 review
Quiz 1 and the Laplace mechanism
Gaussian mechanism and report noisy max
Composition and privacy accounting
No class: Fall Reading Days
DP-SGD intuition and Opacus
Async DP training systems
Async sampler and accounting case study
Exponential mechanism and private selection
Opacus audit studio and JAX step accounting
Private synthetic data foundations
Google DPSynth and release auditing
No class: Election Day
Quiz 2 and PET trust models
MPC, HE, and TEE trade-offs
Network privacy, telemetry, and metadata
Async CCS 2026 paper spotlight
Project break exchange and repair planning
Project repair, transfer, and poster clinic
No class: Thanksgiving recess
Poster / demo session, group 1
Poster / demo session, group 2
Course wrap-up and next steps
No class: finals period
For a maintained collection of courses, books, tutorials, and software, see the OpenDP educational resources.