The Odds are Odd: A Statistical Test for Detecting Adversarial Examples

Explore a statistical approach for detecting adversarially manipulated examples in machine learning models through this 30-minute video presentation. Delve into the research conducted by Kevin Roth, Yannic Kilcher, and Thomas Hofmann, which investigates conditions for reliable test statistics to identify adversarial attacks. Learn about the anomalies introduced by adversarial manipulations, particularly those optimized under p-norm constraints. Discover how these statistics can be computed and calibrated using random input corruption, and understand the requirements for implementing this defense mechanism. Examine the empirical justification for this approach and the conditions that guarantee its effectiveness. Gain insights into the potential for correcting test-time predictions affected by adversarial attacks with high accuracy.