Adversarial Robustness

Global Guarantees of Robustness: A Probabilistic Approach to AI Safety

Rather than certify every point, Mu and Lim estimate the probability that a random input is non-robust, wrapping a Clopper-Pearson interval around a small sample.