Adversarial Examples Against
Deep Neural Network based Steganalysis
Yiwei Zhang
CAS Key Laboratory of
Electromagnetic Space Information
University of Science and Technology
of China
Hefei, Anhui, China
zywvvd@mail.ustc.edu.cn
Weiming Zhang
∗
CAS Key Laboratory of
Electromagnetic Space Information
University of Science and Technology
of China
Hefei, Anhui, China
zhangwm@ustc.edu.cn
Kejiang Chen
CAS Key Laboratory of
Electromagnetic Space Information
University of Science and Technology
of China
Hefei, Anhui, China
chenkj@mail.ustc.edu.cn
Jiayang Liu
CAS Key Laboratory of
Electromagnetic Space Information
University of Science and Technology
of China
Hefei, Anhui, China
1229370169@qq.com
Yujia Liu
CAS Key Laboratory of
Electromagnetic Space Information
University of Science and Technology
of China
Hefei, Anhui, China
yjcaihon@mail.ustc.edu.cn
Nenghai Yu
CAS Key Laboratory of
Electromagnetic Space Information
University of Science and Technology
of China
Hefei, Anhui, China
ynh@ustc.edu.cn
ABSTRACT
Deep neural network based steganalysis has developed rapidly in
recent years, which poses a challenge to the security of steganog-
raphy. However, there is no steganography method that can ef-
fectively resist the neural networks for steganalysis at present. In
this paper, we propose a new strategy that constructs enhanced
covers against neural networks with the technique of adversarial
examples. The enhanced covers and their corresponding stegos are
most likely to be judged as covers by the networks. Besides, we use
both deep neural network based steganalysis and high-dimensional
feature classiers to evaluate the performance of steganography
and propose a new comprehensive security criterion. We also make
a tradeo between the two analysis systems and improve the com-
prehensive security. The eectiveness of the proposed scheme is
veried with the evidence obtained from the experiments on the
BOSSbase using the steganography algorithm of WOW and popular
steganalyzers with rich models and three state-of-the-art neural
networks.
KEYWORDS
Steganography; adversarial examples; deep neural network; ste-
ganalysis; security
∗
Corresponding author
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
IH&MMSec’18, June 20–22, 2018, Innsbruck, Austria
© 2018 Association for Computing Machinery.
ACM ISBN 978-1-4503-5625-1/18/06... $15.00
https://doi.org/10.1145/3206004.3206012
ACM Reference Format:
Yiwei Zhang, Weiming Zhang, Kejiang Chen, Jiayang Liu, Yujia Liu, and Neng-
hai Yu. 2018. Adversarial Examples Against Deep Neural Network based
Steganalysis. In Proceedings of 6th ACM Information Hiding and Multimedia
Security Workshop (IH&MMSec’18). ACM, New York, NY, USA, Article 4,
6 pages. https://doi.org/10.1145/3206004.3206012
1 INTRODUCTION
In recent years, information hiding researchers have proposed many
advanced steganographic algorithms to hide secret information into
a cover image. Most of the schemes embed secret messages in spa-
tial domain or frequency domain, such as HUGO [
16
], WOW [
7
],
S-UNIWARD [
8
], HILL [
14
], J-UNIWARAD [
8
] and UERD [
6
]. These
methods can minimize a heuristically-dened embedding distor-
tion while hiding secrets into a given image to lower the statistical
detectability. And based on an oracle used to calculate the detectabil-
ity map, a new steganography called ASO [
13
] is proposed which
can preserve both cover image and sender’s database distributions
during the embedding process.
In order to detect whether there is hidden information in an
image, the traditional method of steganalysis is divided into two
steps, high-dimensional feature extraction and machine learning
classier training. An excellent steganalyzer is the Rich Model (RM),
which is usually used in the rst step. There are several versions of
Rich Models such as Spacial Rich Model (SRM) [
4
] and its variants
[
3
,
19
] in spatial domain and JPEG-SRM (J-SRM) [
10
] in frequency
domain. The most common choice of machine learning classier is
Ensemble Classier (EC) [
11
]. The combination of SRM and EC has
achieved excellent detection performance.
In the past two years, steganalysis based on Convolutional Neural
Network (CNN) models has made a tremendous progress. Com-
pared with the traditional methods, CNN-based steganalysis uses
various network structures to learn the eective features of images
to distinguish cover images and stego images. Qian [
17
] used a
CNN architecture with Gaussian activations function to construct
Session: Deep Learning for Steganography
IH&MMSec’18, June 20–22, 2018, Innsbruck, Austria