没有合适的资源?快使用搜索试试~ 我知道了~
温馨提示
主要内容为提出了针对攻击具有高鲁棒性的实用可逆神经网络PRIS,其在现有基于可逆神经网络的方法基础上加入了增强模块并提出训练方法以及解决量化误差问题的有效方案,在多项指标上均表现出显著优势。 适用人群包括但不限于从事数据隐藏与提取的研究者和开发人员。 本研究旨在构建一种能够抵御多种实际攻击的数据隐藏方式。适用于需要将敏感信息隐藏到图片载体的应用场景,如版权保护、数字通信等。 研究人员可通过实验对比评估本方法相对于传统方案的效果提升,利用其代码和网页工具快速验证方法的实际应用效果。
资源推荐
资源详情
资源评论
arXiv:2309.13620v2 [cs.CV] 28 Nov 2023
PRIS: Practical robust invertible network for
image steganography
Hang Yang
1
, Yitian Xu
1
* Xuhua Liu
1
* Xiaodong Ma
1
*
1
College of Science, China Agricultural University, Beijing 100083, China
Abstract
Image steganography is a technique of hiding secret information inside another im-
age, so that the secret is not visible to human eyes and can be recovered when
needed. Most of the existing image steganography methods have low hiding robust-
ness when the container images affected by distortion. Such as Gaussian noise and
lossy compression. This paper proposed PRIS to improve the r obustness of image
steganography, it based on invertible neural networks, and put two enhance mod-
ules before and after the extraction process with a 3-step training strategy. Moreover,
rounding error is considered which is always ignored by existing methods, but ac-
tually it is unavoidable in practical. A gradient approximation function (GAF) is
also proposed to overcome the undifferentiable issue of rounding distortion. Exper-
imental results show that our PRI S outperforms the state-of-the-art robust image
steganography method in both robustness and practicability. Codes are available at
https://github.com/yanghangAI/PRIS, demonstration of our model in practical
at http://yanghang.site/hide/.
1 Introduction
Steganography is the art of concealing informatio n. The goal of steganogra-
phy is to hide a secret message within a host medium [1,2]. The host medium
containing the hidden secret message is called the container [3]. The container
is typically publicly visible, but the difference between the host and container
should be invisible to third parties. There are some good introductions to
steganography in [1,4,5]. Image steganography aims to hide message such as
image, audio, and text within a host image in an undetectable way [6], mean-
ing the host medium in image steganography is an image. Nowadays, image
Email address: xytshuxue@126.com, Tel.: +8610 62737077. (Yitian Xu
1
*).
steganography is widely used in fields such as copyright protection, digital
communication, information certification, and more [7].
Traditional image steganographic methods have the ability to hide only a small
amount of information [8,9 ,1 0,11,12,13]. They conceal secret messages in the
spatial or adaptive domains [14] with capacities of 0.2∼4 bits per pixel (bpp).
In recent years, some researchers have attempted to hide an image conta ining
more information than the secret message in traditional image steganographic
methods within a host image [15,16,17,18,19,20]. They introduced deep learn-
ing into image steganography by using two separate networks to realize the
embedding and extraction processes. Lu [3] and Jing [21] achieved state-of-
the-art performance in image steganography by using an invertible network
to embed a secret image into a host image. Since the extraction process is the
inverse of the embedding process, the secret image can be fully recovered.
In practical applications, the container image is of ten subject to various forms
of attacks due to factors such as lossy image compression for storage savings.
These attacks can distort the container image, potentially affecting the ability
to extract the secret information. Robustness refers to the ability of the secret
information to remain unchanged when the container is distorted [22]. In other
words, it measures how similar the secret and extracted messages will be when
the container image is attacked.
However, state-of-the-art image steganography methods [3,21] did not take
into account the effect of container images being attacked. As a result, they
failed to extract secret images when container images were distorted. To ad-
dress this issue, Xu [6] proposed robust inver t ible image steganography. This
method improved the robustness of image steganography by taking image
distortion into account, allowing for the successful extraction of secret infor-
mation even when the container image is distorted.
Moreover, since the current mainstream deep learning frameworks have a nu-
merical precision of 32 bits, while images are usually 8 bits, there is an in-
evitable rounding error when saving the container images. However, to the
best of our knowledge, the existing deep steganography methods ignore this
problem [23,6,21,3], which is not practical in real application, we also give a
method to illustrate the importance to consider rounding error in section 3.7.
In this paper, we designed a practical robust invertible network called PRIS
based on HiNet [21]. We introduce two enhancement mo dules and a 3-step
training strategy into invertible neural network to improve its robustness. In
addition, we take rounding distortion into account and propose a g r adient
approximation f unction (GAF ) to deal with the undifferentiable problem of
rounding operation. The main contributions are listed as follows:
1. A practical robust invertible network is proposed for image steganography
2
under diverse a t tacks.
2. We introduce a 3-step training strategy into our training process to achieve
a better ro bustness.
3. A gradient approximate function is proposed to solve the undifferentiable
problem caused by rounding operation, and take rounding error into consider-
ation.
4. Experiments results demonstrate that our proposed PRIS outperforms the
existing state-of-the-art method RIIS in both robustness and practicability.
2 Related work
2.1 Traditional image steganography
Traditional image steganography can be divided into two categor ies based on
the whether the hiding process happen in spatial or frequency domain. Spatial
domain: The most popular spatial-based method called Least Significant Bit
(LSB) [24,25], it changes n least significant bits of host image to embed secret
messages. However, the change of picked bits make it easy t o be detected by
some steganalysis methods [9,26,27]. In addition, Pan [28] utilizes pixel value
differencing (PVD), Tsai [29] introduces histogram shifting, Nguyen [30] use
multiplebit-planes and Imaizumi [31] propose palettes in image steganography.
Frequency domain: Those methods hide secret messages in frequency do-
mains, such as discrete cosine transform (DCT) [10], discrete Fourier transform
(DFT) [13], and discrete wavelet transform (DWT) [8] domains. Although they
are more robust and undetectable than LSB methods, they still suffered from
limited payload capacity.
2.2 Deep learning-based image steganography
Recently, many researchers have a pplied deep learning methods to image
steganography and achieved better performance than traditional methods.
HiDDeN [32] and SteganoGAN [33] utilize the encoder-decoder architecture to
realize the embedding and extraction of secret messages and employ a third
network t o resist steganalysis adversarially. Shi [34] proposes Ssgan for im-
age steganography, which is ba sed on generative adversarial networks. Baluja
[17,15] and Zhang [16] hide a secret image of the same size as the host im-
age with deep learning method, achieve a much higher payload capacity than
traditional methods.
3
2.3 Invertible neural network
Dinh [35] proposed Invertible neural networ k (INN) in 2014. It learns a stable
bijective mapping between a data distribution p
X
and a latent distribution p
Z
.
Unlike CycleGAN [36], which uses two generators and a cycle loss to achieve
bidirectional mapping, INN performs both f orward and backward propagation
within the same networ k, acting as both feature encoder and image generator.
INNs are also useful for estimating t he posterior of an inverse problem [37].
More flexible INNs are built with masked convo lutions under some composi-
tion rules in [38]. An unbiased flow-based generative model is proposed in [39].
Other works improve the coupling layer for density estimation, such as Glow
[40] and i-ResNet [41 ], resulting in better generation quality.
Lu [3], Jing [21] and Jia [42] introduced INN into image steganography. The
strict invertibility of INN just meets the requirement that embedding and
extraction are mutually inverse processes. Therefore they gain state-of-the-
art performance in image steganography. However, the strict invertibility also
means that when the container image is attacked, the noise is also transmitted
to the extracted image, which make it vulnerable to attack. Xu [6] proposed
robust invertible image steganography (RIIS) and improve the robustness of
the afor ementioned invertible neural netwo rks.
In the field of image steganography, the invertible network has shown its great
potential. It achieves state-of-the-art performance by utilizing t he prior knowl-
edge that hiding and extraction are inverse processes [3,21], along with the
strict reversibility of its own network. However, due to its strictly reversible
nature, when the container image is perturbed, t he inverse process of its net-
work will also be perturbed. This implies that the extracted image will also be
distorted. To address this issue, we proposed our model PRIS, by introducing
two new modules: pre-enhance and post-enhance, which are added before and
after the extraction process respectively, and through a 3-step training strat-
egy, we improve its robustness to state-of-the-art level. Moreover, rounding
error is considered in our PRIS since it is unavoidable in practice, and a gradi-
ent approximate function is proposed to address the undifferentiable problem
of rounding operation.
4
剩余18页未读,继续阅读
资源评论
pk_xz123456
- 粉丝: 2581
- 资源: 3641
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
- 超好看倒计时特效单页html模板源码.zip
- 荒地、农田、森林、湖、山姆、住宅检测11-YOLO(v5至v9)、COCO、CreateML、Darknet、Paligemma、TFRecord、VOC数据集合集.rar
- 基于epoll的reactor模型
- 人力资源领域人员简历模板docx文档
- 使用python基于CNN的10种水果识别-含1w张以上的数据集图片
- 基于Delaunay三角化的点云数据三维曲面重建matlab仿真,包括程序,中文注释,仿真操作步骤视频
- 船舶检测20-YOLO(v5至v9)、COCO、CreateML、Darknet、Paligemma、TFRecord、VOC数据集合集.rar
- 船舶检测19-YOLO(v5至v9)、COCO、CreateML、Darknet、Paligemma、TFRecord、VOC数据集合集.rar
- 华为ENSP基本配置!!!
- Java高级软件工程师简历模板-技能特长与项目经历
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功