Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks

TL;DR

This paper introduces feature squeezing, a simple and efficient method to detect adversarial examples in deep neural networks by reducing input complexity and comparing model predictions.

Contribution

The paper proposes feature squeezing as a novel, computationally inexpensive defense mechanism for identifying adversarial examples in DNNs.

Findings

High detection accuracy for adversarial examples

Effective when combined with other defense strategies

Works with simple input transformations like color depth reduction and smoothing

Abstract

Although deep neural networks (DNNs) have achieved great success in many tasks, they can often be fooled by \emph{adversarial examples} that are generated by adding small but purposeful distortions to natural examples. Previous studies to defend against adversarial examples mostly focused on refining the DNN models, but have either shown limited success or required expensive computation. We propose a new strategy, \emph{feature squeezing}, that can be used to harden DNN models by detecting adversarial examples. Feature squeezing reduces the search space available to an adversary by coalescing samples that correspond to many different feature vectors in the original space into a single sample. By comparing a DNN model's prediction on the original input with that on squeezed inputs, feature squeezing detects adversarial examples with high accuracy and few false positives. This paper explores two feature squeezing methods: reducing the color bit depth of each pixel and spatial smoothing. These simple strategies are inexpensive and complementary to other defenses, and can be combined in a joint detection framework to achieve high detection rates against state-of-the-art attacks.

Figures

Full text

See pages - of NDSS18_Feature_Squeezing.pdf