Differentially Private Statistical Estimation, Machine Learning, and Testing
Stub. Applying differential privacy to real statistical workloads: private deep learning via DP-SGD (Abadi et al.), differentially private statistical estimation and PAC learning (Gaboardi et al.), and private algorithms for hypothesis testing and data analysis (McSherry-Talwar).
Corresponds to CS/ECE 598DA, Week 5. Moves from DP’s basic mechanisms to the engineering of private statistical procedures at scale. DP-SGD (Abadi et al.) is the dominant approach to private deep learning: clip per-example gradients to bound sensitivity, add noise, and track the accumulated privacy loss across many training steps via composition. Gaboardi et al. covers differentially private statistical estimation and learning-theoretic (PAC-style) guarantees for private learners, and McSherry-Talwar covers private mechanisms for auctions/mechanism design and hypothesis testing. See machine-learning-and-data-privacy for the broader ML-security/privacy context (membership inference, model inversion) this sits alongside.
Outline (TODO — flesh out each)
- DP-SGD — private deep learning (Abadi et al.)
- Privacy budget accounting across training steps
- Differentially private statistical estimation and learning (Gaboardi et al.)
- Private mechanism design and hypothesis testing (McSherry-Talwar)