Defensive Distillation is Not Robust to Adversarial Examples

TL;DR

This paper demonstrates that defensive distillation, previously thought to improve neural network robustness, does not actually provide resistance against targeted adversarial attacks.

Contribution

The study reveals that defensive distillation fails to enhance neural network security against adversarial examples, challenging prior assumptions.

Findings

Defensive distillation does not improve robustness against targeted attacks.

Neural networks with defensive distillation are as vulnerable as unprotected models.

The paper provides evidence that defensive distillation is ineffective for security.

Abstract

We show that defensive distillation is not secure: it is no more resistant to targeted misclassification attacks than unprotected neural networks.