NSF CAREER: Exact Optimal and Data-Adaptive Algorithms and Tools for Differential Privacy |
|
Principal Investigator Yu-Xiang Wang, UCSB (2021 - 2024), UCSD (2024 -) |
Project Summary Talks | Products | Education |
Funded by NSF CNS 2048091.
This material is based upon work supported by the National
Science Foundation under Grant No. 2048091. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of the
author(s) and do not necessarily reflect the views of the National Science
Foundation.
Differential privacy (DP) is a rigorous way to learn from sensitive data while limiting what can be inferred about any one person. This CAREER project developed mathematical foundations, algorithms, software, and educational resources that make DP more accurate, adaptable, and practical. The research produced exact tools for tracking privacy risk across complex analyses; data-adaptive and per-person privacy methods; and private learning techniques for classification, data valuation, reinforcement learning, online decision-making, and language models. The project also expanded the open-source AutoDP library so that researchers and practitioners can use state-of-the-art privacy accounting more reliably.
The award supported doctoral students, a postdoctoral researcher, and undergraduate researchers, and its results were incorporated into graduate teaching and hands-on student projects. Recent work extends the project to the responsible use of artificial intelligence: Private-RAG protects document collections while answering many questions; DPCheatSheet teaches people to recognize common privacy-programming errors; and DPrivBench tests whether large language models can reason correctly about formal privacy guarantees. Papers, code, tutorials, and presentation materials are publicly available below.
Fig. 1. Exact optimal DP accounting represents a privacy mechanism by a function and converts losslessly among commonly used privacy descriptions.
Fig. 2. Project overview presented at the 2024 NSF Secure and Trustworthy Cyberspace Principal Investigator Meeting. Click the poster to view the PDF.
- Oral presentation on "DPrivBench: Benchmarking LLMs' Reasoning for Differential Privacy"
Erchi Wang presented at the Theory and Practice of Differential Privacy (TPDP) 2026 workshop. - Oral presentations on "Adapting to Linear Separable Subsets with Large-Margin in Differentially Private Learning" and "Purifying Approximate Differential Privacy with Randomized Post-processing"
Erchi Wang and Yingyu Lin presented at the Theory and Practice of Differential Privacy (TPDP) 2025 workshop. - Invited presentation on recent advances in differential privacy
PI Wang presented at a differential privacy workshop in Bertinoro, Italy, in 2025. - Conference presentations of private learning and privacy-purification results
PI Wang and coauthors presented project-supported work at ICML 2025 and NeurIPS 2025; "Purifying Approximate Differential Privacy with Randomized Post-processing" received a NeurIPS Spotlight. - Project poster at the 2024 NSF SaTC Principal Investigator Meeting
PI Wang presented the project's research, software, and educational outcomes in Pittsburgh, Pennsylvania. [Poster PDF] - Presentations and AutoDP dissemination at TPDP 2024
PI Wang and coauthors presented project-supported research and shared AutoDP with the community in Boston. - Paper "Neural Collapse Meets Differential Privacy" is selected for an Oral Presentation at ICML 2024
Among the top 0.3% of all submissions. - Invited talk on "Open Problem: Near Linear time algorithm for DP-ERM"
PI Wang spoke at TTIC Federated Learning Workshop, 2023 - Invited talk on "How to Squeeze out the Last Drop of Utility in your DP application"
PI Wang spoke at TTIC Federated Learning Workshop, 2023 - Invited talk on "Towards Practical Differential Privacy"
PI Wang spoke at Nokia Bell Lab, UCSD, Princeton in 2023 - Keynote talk on "Practical Differential Privacy"
PI Wang spoke at NeurIPS'22 Workshop on "Challenges in Deploying and Monitoring Machine Learning Systems" - Invited talk on "Towards Practical Differential Privacy with Differential Privacy-like ideas"
PI Wang spoke at Google Research Workshop on "Security, Privacy, and Anti-Abuse Research Workshop" - Invited talk on "Optimal Accounting of Differential Privacy"
PI Wang spoke at MIT Media Lab and LinkedIn Data Science Seminar Series - Invited talk on "Optimal Accounting of Differential Privacy via Characteristic Function"
PI Wang spoke at Rutgers. Jinshuo and Yuqing spoke at Google. - Invited talk on "Per-instance Differential Privacy and How to Privately Publish Them"
PI Wang spoke at Google and at University of Albany - Invited talk on "Advances in Differential Privacy and Private Federated Learning"
PI Wang spoke at Amazon Web Services. - News article "A Matter of Privacy"
Coverage on this project on UCSB Magazine "The Current" [Link] - Invited talk on "Privacy Amplification by Sampling and Renyi Differential Privacy"
PI Wang spoke at Berkeley Simons Institute [Link to the talk]
- DPrivBench: Benchmarking LLMs' Reasoning for Differential Privacy
Manuscript. [arxiv] - Private-RAG: Answering Multiple Queries with LLMs while Keeping Your Data Private
COLM 2026. [arxiv] - Improved Regret in Stochastic Decision-Theoretic Online Learning under Differential Privacy
ALT 2026. [arxiv] - Purifying Approximate Differential Privacy with Randomized Post-processing
NeurIPS 2025. (*Spotlight Presentation) [arxiv] - DPCheatSheet: Using Worked and Erroneous LLM-usage Examples to Scaffold Differential Privacy Implementation
Manuscript under submission. [arxiv] - Adapting to Linear Separable Subsets with Large-Margin in Differentially Private Learning
ICML 2025. [arxiv] - Differentially Private Reinforcement Learning with Self-Play
NeurIPS 2024. [arxiv] - Differentially Private Bias-Term only Fine-tuning of Foundation Models
ICML 2024. [arxiv] - Neural Collapse meets Differential Privacy: Curious behaviors of NoisyGD with Near-Perfect Representation Learning
ICML 2024. [arxiv] (*Oral Presentation) -
Privacy Profiles for Private Selection
ICML 2024. [arxiv] -
Tractable MCMC for Private Learning with Pure and Gaussian Differential Privacy
ICLR 2024. [arxiv] - Advancing Differential Privacy: Where We Are Now and Future Directions for Real-World Deployment
Harvard Data Science Review [paper, arxiv] -
Improving the Privacy and Practicality of Objective Perturbation for Differentially Private Linear Learners
NeurIPS 2023. [arxiv] - Threshold KNN-Shapley: A Linear-Time and Privacy-Friendly Approach to Data Valuation
NeurIPS 2023. [arxiv] (*Spotlight Presentation) - Offline Reinforcement Learning with Differential Privacy
NeurIPS 2023. [arxiv] - "Private Prediction Strikes back!" Private Kernelized Nearest Neighbors with Individual Renyi Filter
UAI 2023. [arxiv] (*Spotlight Presentation) - Near-Optimal Differentially Private Reinforcement Learning
AISTATS 2023 [arxiv] - Generalized PTR: User-Friendly Recipes for Data-Adaptive Algorithms with Differential Privacy
AISTATS 2023. [arxiv] (*Plenary Oral Presentation) - Differentially Private Linear Sketches: Efficient Implementations and Applications
NeurIPS 2022. [arxiv] - SeqPATE: Differentially Private Text Generation via Knowledge Distillation
NeurIPS 2022. [openreview] - Provably Confidential Language Modelling
NAACL 2022. [arxiv] (*Oral Presentation) - Adaptive Private-K-Selection with Adaptive K and Application to Multi-label PATE
AISTATS 2022. [arxiv] - Optimal Accounting of Differential Privacy via Characteristic Function
AISTATS 2022. [arxiv] - Privately Publishable Per-instance Privacy
NeurIPS 2021. [arxiv] -
Revisiting Model-Agnostic Private Learning: Faster Rates and Active Learning
Journal of Machine Learning Research. Shorter version appeared in AISTATS 2021. [arxiv] - Voting-based Approaches For Differentially Private Federated Learning
Manuscript. [arxiv] - Subsampled Renyi Differential Privacy and Analytical Moments Accountant
Journal of Privacy and Confidentiality, 2021 [ paper ]. A shorter version appeared at AISTATS 2019 and received a Notable Paper Award.
- DSC 291: Differential Privacy—Theory, Algorithms and Applications [Course website]
Instructor: Yu-Xiang Wang, UC San Diego, Fall 2024. - Privacy in Practice: hands-on privacy telemetry projects [Project website]
Undergraduate capstone projects connecting formal privacy guarantees with practical data collection and machine-learning applications. - CS291A Introduction to Differential Privacy: Theory, Algorithms and Applications [Course website]
Instructor: Yu-Xiang Wang, 2021 Fall - [Open source project] New API / examples / tutorials / unit tests for
autodp
Contributors: Yu-Xiang Wang, Yuqing Zhu, Borja Balle (DeepMind), Stefan Mallem (Google) - Guest Lecture on "Composition, RDP and Amplification by Sampling"
at Mijung Park's CPSC 532P at University of British Columbia [Course website]
Fig. 3. The AutoDP application programming interface for composing and calibrating differentially private mechanisms.