[1]赵俊杰,王金伟.基于SmsGAN的对抗样本修复[J].郑州大学学报(工学版),2021,42(01):50-55.[doi:10.13705/j.issn.1671-6833.2021.01.008]
　Zhao Junjie,Wang Jinwei,Recovery of Adversarial Examples ba＜x＞sed on SmsGAN[J].Journal of Zhengzhou University (Engineering Science),2021,42(01):50-55.[doi:10.13705/j.issn.1671-6833.2021.01.008]

点击复制

基于SmsGAN的对抗样本修复()

分享到：

《郑州大学学报(工学版)》[ISSN:1671-6833/CN:41-1339/T]

卷:

42卷

期数:

2021年01期

页码:

50-55

栏目:

出版日期:

2021-03-14

文章信息/Info

Title:

Recovery of Adversarial Examples ba＜x＞sed on SmsGAN

作者:

赵俊杰; 王金伟

南京信息工程大学计算机与软件学院;中国科学院信息工程研究所信息安全国家重点实验室;

Author(s):

Zhao Junjie; Wang Jinwei;

School of Computer and Software, Nanjing University of Information Engineering; National Key Laboratory of Information Security, Institute of Information Security, Institute of Information Engineering, Institute of Information Engineering;

关键词:

深度学习; 对抗样本; 图像取证; 样本修复; 生成对抗网络

Keywords:

Deep learning;  confrontation sample;  image evidence;  sample repair;  generate confrontation network

DOI:

10.13705/j.issn.1671-6833.2021.01.008

文献标志码:

A

摘要:

对抗样本攻击对于深度卷积神经网络是一个极大的威胁，然而对抗样本本身具有脆弱性，使得其修复成为可能。随机多滤波特征统计生成对抗网络（SmsGAN）以随机多滤波特征统计网络（SmsNet）为判别器，并采用目标引导生成器。在SmsNet中，我们设计了用于获取特征图全局特性的特征统计层，并将每个卷积层输出的特征图直接送到特征统计层，从而实现了对抗样本的高精确度取证。生成器采用多尺度卷积核并行结构避免棋盘效应的产生，损失函数由判别损失和引导损失两部分组成，形成目标引导生成器。对抗样本经过下采样网络获取局部统计特征，再输入SmsGAN得到修复的样本。实验表明，采用SmsGAN修复对抗样本，在保证修复效果的同时可以保持高视觉质量。

Abstract:

The attack of adversarial examples is a great threat to the deep convolutional neural network. However, the vulnerability of adversarial examples makes it possible to repair them. Stochastic multifilter statistical generative adversarial network (SmsGAN) uses the Stochastic multifilter statistical network (SmsNet) as its discriminator and adopts a target guidance generator. In SmsNet, we design a feature statistical la＜x＞yer to obtain the global characteristics of the feature maps and send the feature maps output by each convolution la＜x＞yer directly to the feature statistical la＜x＞yer. Then high-precision forensics of adversarial examples are achieved. The generator uses a multi-scale convolution kernel parallel structure to avoid the occurrence of the checkerboard artifacts. The loss function consists of two parts: discriminant loss and guide loss to form the target guide generator. The loss function of the generator consists of two parts: discriminative loss and guidance loss, to form a target guidance generator. The adversarial examples enter into the down-sampling network to obtain local statistical features, and then these features are sent into SmsGAN for reconstruction to get denoised examples. Experiments show that using SmsGAN to recover adversarial samples can maintain high visual quality while ensuring the denoising effect.

常用功能