Explore the privacy implications of machine unlearning in this 50-minute Google TechTalk presented by Mohammad Mahmoody as part of the Differential Privacy for ML seminar series. Delve into the machine learning framework, threat models, and security games associated with deleting inferences and membership inference. Examine theoretical intuitions behind label memorization and reconstruction thread models. Analyze experimental results and discuss deletion compliance in the context of machine learning privacy concerns.
Machine Unlearning and Privacy Implications - Differential Privacy for ML
Overview
Syllabus
Introduction
What is this talk about
Machine Learning Framework
Hypothesis A
Outline
Threat Models
Deleting Inference
Security Game
Membership Inference
Meta Attacks
Theoretical Intuition
Label Memorization
Experiment
Reconstruction
Thread Model
Experimental Results
Deleting Compliance
Taught by
Google TechTalks
Related Courses
-
Auditing Data Privacy for Machine Learning
-
Control, Confidentiality, and the Right to be Forgotten in Differential Privacy for Machine Learning
-
Generative Adversarial Models for Privacy and Fairness
-
Membership Inference Attacks against Machine Learning Models
-
Adversary Instantiation - Lower Bounds for Differentially Private Machine Learning
