"SECRETS ARE LIES, SHARING IS CARING, PRIVACY IS THEFT." - A Dive into Privacy Preserving
summary
Describes concrete attacks and defenses rather than slogans: shows linkage, reconstruction, model inversion, membership inference attacks including a demo that builds shadow models and recovers membership with high accuracy on a financial purchases dataset, then applies DP‑SGD optimizer from TensorFlow Privacy and PATE-style voting to inject noise. Reports that membership accuracy drops substantially under differential privacy at the cost of lower test accuracy, and recommends combining HE, MPC or federated learning with DP for practical deployments.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:46Motivation
- 1:21What is Privacy?
- 1:39Privacy vs Utility Tradeoff
- 2:25The Attack Surface
- 3:00Linkage Attack
- 4:31De-anonymization Attack (Netflix Example)
- 5:15Reconstruction Attack
- 6:01Model Inversion Attack
- 6:30Membership Inference Attack Train Attack Model using Shadow Models
- 6:57Query Attack
- 7:37Privacy Attack Comparison
- 7:57Privacy Preserving ML
- 8:25Privacy Aware Machine Learning Secure computation
- 8:51Homomorphic Encryption
- 9:35Secure MultiParty Computation
- 10:19Federated Machine Learning
- 10:52Differential Privacy
- 13:46Private Aggregation of Teacher Ensembels
- 15:31Defense method comparison
- 15:46Combinatory Methods
related talks
01h 17mTutorial: Privacy, Part I
33mPrivacy of Decentralized Machine Learning
28mPrivacy preserving machine learning
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
14mEstablishing Privacy Metrics for Genomic Data Analysis
38mDemystifying Privacy Preserving Computing
20mDesigning for User Privacy: Integrating Differential Privacy into Ad Measurement...
29mKeynote Presentation: Minding the (Privacy Research/Practice) Gap