Adversarial and Clean Data Are Not Twins

TL;DR

This paper demonstrates that adversarial and clean data can be effectively separated with high accuracy using a simple classifier, revealing fundamental differences and limitations in current defense methods.

Contribution

The paper introduces a simple binary classifier that reliably distinguishes adversarial from clean data and explores the intrinsic limitations of existing adversarial defenses.

Findings

Binary classifier achieves over 99% accuracy in separating adversarial from clean data.

The classifier remains robust against second-round adversarial attacks.

Current defense methods are limited by intrinsic properties of adversarial crafting algorithms.

Abstract

Adversarial attack has cast a shadow on the massive success of deep neural networks. Despite being almost visually identical to the clean data, the adversarial images can fool deep neural networks into wrong predictions with very high confidence. In this paper, however, we show that we can build a simple binary classifier separating the adversarial apart from the clean data with accuracy over 99%. We also empirically show that the binary classifier is robust to a second-round adversarial attack. In other words, it is difficult to disguise adversarial samples to bypass the binary classifier. Further more, we empirically investigate the generalization limitation which lingers on all current defensive methods, including the binary classifier approach. And we hypothesize that this is the result of intrinsic property of adversarial crafting algorithms.