AdvIBP: Certified Adversarial Training by Combining Adversarial Training and Provable Robustness Verification in a Principled Way

  • We achieved state-of-the-art verified (certified) error on MNIST and CIFAR: for MNIST, 6.60% at epsilon=0.3 and 12.30% at epsilon=0.4 (for L_infinity norm perturbations); and for CIFAR, 66.57% at epsilon=8/255 and 76.05% at epsilon=16/255 (also for L_infinity norm perturbations).
  • Github link:
  • Related publication(s):
    1. Jiameng Fan and Wenchao Li. Adversarial Training and Provable Robustness: A Tale of Two Objectives. The 35th AAAI Conference on Artificial Intelligence (AAAI), February 2021. 

ReachNN*: Reachability Analysis Tool of Neural Network Controlled Systems (NNCSs)