Adversarial Examples Detection and Analysis with Layer-wise Autoencoders

TL;DR

This paper introduces a layer-wise autoencoder approach to detect adversarial examples by analyzing data representations within neural network layers, providing improved detection performance and insights into adversarial behavior.

Contribution

The paper proposes a novel method using autoencoders at intermediate layers to identify adversarial examples and analyze their flow through neural networks.

Findings

Outperforms existing methods in detection accuracy

Provides insights into adversarial example propagation

Effective in both supervised and unsupervised scenarios

Abstract

We present a mechanism for detecting adversarial examples based on data representations taken from the hidden layers of the target network. For this purpose, we train individual autoencoders at intermediate layers of the target network. This allows us to describe the manifold of true data and, in consequence, decide whether a given example has the same characteristics as true data. It also gives us insight into the behavior of adversarial examples and their flow through the layers of a deep neural network. Experimental results show that our method outperforms the state of the art in supervised and unsupervised settings.