Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers

Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers

Microsoft: Hadi Salman, Greg Yang, Jerry Li, Pengchuan Zhang, Huan Zhang, Ilya Razenshteyn, Sébastien Bubeck

Contribution:

• employ adversarial training to improve the performance of randomized smoothing.
• state-of-the-art results for l2 norm
• a more concise proof of tight robustness guarantee by casting this as a non-linear Lipschitz property

Not a new certification method, the improvement are due to the better base classifiers as a result of adversarial training

Author’s blog: https://decentdescent.org/smoothadv.html

Author’s video: Code: https://github.com/Hadisalman/smoothing-adversarial

---

• Previous
  Certified Adversarial Robustness via Randomized Smoothing
• Next
  Attacks Which Do Not Kill Training Make Adversarial Learning Stronger

---