Defense Against the Dark Arts

Explore a comprehensive keynote address on adversarial example research in machine learning and cybersecurity. Delve into the intricacies of defending against attacks on the machine learning pipeline, including transfer attacks, gradient masking, and norm ball scenarios. Examine the concept of adversarial logit pairing (ALP) and investigate future research directions in indirect methods and improved attack models. Gain insights into non-security applications of adversarial examples and the fascinating "Clever Hans" phenomenon. Learn from Ian Goodfellow's expertise as he presents at the 1st Deep Learning and Security Workshop during the 2018 IEEE Symposium on Security & Privacy in San Francisco.

Intro
I.I.D. Machine Learning
Attacks on the machine learning pipeline
Define a game
Fifty Shades of Gray Box Attacks
Transfer Attack
Norm Balls: A Toy Game
Tradeoff
Gradient Masking
Pipeline of Defense Failures
Adversarial Logit Pairing (ALP)
Future Directions: Indirect Methods
Future Directions: Better Attack Models
Some Non-Security Reasons to Study Adversarial Examples
Clever Hans