On the Effectiveness of Defensive Distillation

TL;DR

This paper demonstrates that defensive distillation effectively reduces the success of various adversarial attacks on neural networks, including both fast gradient sign and Jacobian-based iterative methods.

Contribution

It provides experimental evidence that defensive distillation is a robust defense mechanism against multiple adversarial attack techniques.

Findings

Defensive distillation mitigates fast gradient sign attacks.

It also defends against Jacobian-based iterative attacks.

The method shows broad effectiveness across attack types.

Abstract

We report experimental results indicating that defensive distillation successfully mitigates adversarial samples crafted using the fast gradient sign method, in addition to those crafted using the Jacobian-based iterative attack on which the defense mechanism was originally evaluated.