AI networks are more vulnerable to malicious attacks than previously thought
Illustration of the proposed QuadAttackK method in comparison with the prior art (e.g., the adversarial distillation (AD) method [Zhang and Wu, 2020]). Credit: Thomas Paniagua et al, QuadAttacK: A Quadratic Programming Approach to Learning Ordered Top-K Adversarial Attacks. https://openreview.net/pdf?id=t3vPEjgNtj Artificial...