科研速记(9)：CVPR18-ADV_DEFENCE-Guided Denoiser

---

layout: post
title: CVPR18-ADV_DEFENCE-Guided Denoiser
date: 2019-12-13 16:63:24.000000000 +09:00

PAPER INFO

Defense Against Adversarial Attacks Using High-Level Representation Guided Denoiser
Source Code

key words: adversarial defence, guided denoiser, change loss function, U-net, ladder network

OVERVIEW

What’s the main concern of this paper?

Statement one:

Using denoiser to defend which is to constraint the difference between the l-th representations of target model (e.g., ResNet) activated by clean data $x$ and denoised data $\hat{x}$.

Statement two:

defend by change loss function in order to bypass the amplification effect of adversarial noise.

What’s the main idea/contribution?

Net architecture

Model Name	Intro	Key Words
DUNET (denoising U-net)	modify DAE with U-net and proposed the DUNET	ladder net, lateral connection, residual learning

Denoiser

Model Name	Intro	Type
PGD (pixel guided denoiser)	amplification effect; loss function is defined at the level of image pixels	unsupervised
HGD (high-level representation guided denoiser)	alleviate amplification effect, loss function is defined at the level of** high-level layer’s feature maps**	unsupervised
CGD (class label guided denoiser)	use classification loss function as denoising loss function	supervised

HGD

Model Name	Intro	Key Words
FGD (feature guided denoiser)	also known as perceptual loss or feature matching loss	l=-2, the direct output of** the final conv_layer**
LGD (logits guided denoiser)		l=-1, the output of the layer before the final softmax function

How does it perform(Experiment)?

Anything improvable or eye-opening?

DETAILS