Explore a 28-minute conference talk from USENIX Security '16 that delves into the vulnerabilities of machine learning models deployed with public query interfaces. Learn about model extraction attacks where adversaries aim to duplicate confidential ML models using only black-box access. Discover simple yet efficient techniques for extracting logistic regression, neural network, and decision tree models with near-perfect fidelity. Examine real-world demonstrations against BigML and Amazon Machine Learning services. Investigate potential countermeasures and their limitations, including the impact of omitting confidence values from model outputs. Gain insights into the broader implications for ML model deployment and the need for robust protection strategies in the growing field of ML-as-a-service.
Overview
Syllabus
Intro
Machine Learning (ML) Systems
Model Extraction Attacks (Prior Work)
Main Results
Model Extraction Example: Logistic Regression
Generic Equation Solving Attacks
Online Attack AWS Machine Learning
Application: Model-Inversion Attacks Infer training data from trained models Fredrikson et al.- 2015
Extracting a Decision Tree
Countermeasures
Brief Announcement
Conclusion
Generic Model Retraining Attacks
Taught by
USENIX
Related Courses
-
Intermediate Data Manipulation and Machine Learning
-
Duplicating Black Box Machine Learning Models
-
Are Your Sensitive Attributes Private? Novel Model Inversion Attribute Inference Attacks on Classification Models
-
A Beginner's Guide to Adversarial Machine Learning
-
Fighting New Era Bandits Who Are Plundering Your AI - ML Application
-
Manipulating Machine Learning - Poisoning Attacks and Countermeasures for Regression Learning
