Towards Privacy-Respectful Cyber-Physical Systems: Two Perspectives
Talkat Data Privacy: From Foundations to Applications (Simons Institute workshop, 2019) · 2019-03-04..03-08·
summary
Describes two concrete approaches: a data-valuation pipeline that assigns payments using Shapley value for ML, with an exact k-NN scheme that computes values in O(log n) per test, a sampling estimator needing O(n log n) retrainings, and faster variants (compressive sensing gives O(log log n) under sparsity and influence‑function heuristics) tested on 1,000+ examples and deployed in a Stanford Hospital marketplace trial; and an inferential privacy formulation for smart buildings using mutual information, converted to a linear program to minimize privacy loss under control-performance constraints.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:45Examples of CPS
- 1:40Activity Inference from Energy Data
- 2:17Challenges
- 5:07Machine learning as a Coalitional Game
- 7:18Definition & Properties
- 11:49Opportunities
- 13:12Exact Shapley Value for KNN Classifiers
- 13:52KNN Shapley Values in a Large Dataset
- 14:32Compute the Shapley Value for General Utility Functions • Sampling-based Approximation
- 19:13How to Better Make Use of a Single Model Evaluation?
- 23:08Sample-Efficient Shapley Value Estimation
- 24:50Uniform Value Division for Stable Learning Algorithms • A leaming algorithm is stable if the model learned by the algorithm is insensitive to the removal of an arbitrary point in the
- 25:43Exploiting Influence Functions
- 27:25Compare the Approximation Methods
- 28:42How are the values distributed?
- 29:18Summary
- 31:14Location privacy issues in smart buildings
- 31:58Inferential Privacy
- 33:22Comparison with other privacy mechanisms
- 35:51Duet: Automatic Differential Privacy Proofs
related talks
31mScaling privacy embedded distributed ML systems
29mPrivacy-preserving methods: Building secure projects
42mPrivacy Challenges in Machine Learning
38mDemystifying Privacy Preserving Computing
14mEstablishing Privacy Metrics for Genomic Data Analysis
01h 19mWhen Data Can’t Move but Insight Must
01h 20mUnlocking societal benefits of AI with privacy protective technologies.
28mPrivacy Engineering Networking Break AMA