Overview
Explore a conference talk on certified robustness to adversarial examples using differential privacy. Delve into the innovative PixelDP defense, which scales to large networks and datasets while providing guarantees against norm-bounded attacks. Learn about the connection between adversarial robustness and differential privacy, and understand how this approach offers a rigorous, generic, and flexible foundation for defending machine learning models, particularly deep neural networks. Examine the application of this defense to large-scale networks like Google's Inception for ImageNet, and gain insights into key questions, challenges, and results related to this cutting-edge security research.
Syllabus
Intro
Adversarial Examples
ImageNet Network
Key Questions
Core Problem
Differential Privacy
Summary
Challenges
Scale
Autoencoder
Postprocessing
Guarantees
Results
Robustness threshold
Summarize
Questions
Taught by
IEEE Symposium on Security and Privacy