Abstract

Nowadays, while steganography is the main mean of illegal secret communication, the need of detecting steganographic content and especially stego images is becoming more compulsory. Since multimedia content can be easily spread over the internet and more complicated steganography algorithms in different domains i.e. spatial, transform are utilized, the task of identifying stego images becomes very difficult. Early steganalysis methods deploy statistical attacks on stego images while more recent ones use deep learning techniques. The latter ones mainly utilize convolutional neural networks and show promising results. In this paper we propose a novel method to identify stego images derived from two different steganographic algorithms S-UNIWARD (Spatial-UNIversal WAvelet Relative Distortion) and WOW (Wavelet Obtained Weights) for various embedding rates. The proposed method initially utilizes a dilated convolutional neural network as a feature extractor and afterwards the extracted feature vector trains a random forest classifier. More specifically it is proved that in steganalysis, a dilated convolutional neural network could be an excellent feature extractor and the traditional softmax layer could be replaced by another machine learning classifier. Extensive experiments were conducted, and the proposed model was also compared against state-of the-art convolutional neural networks utilized in spatial image steganalysis, and other feature extraction methods. Results showed that the proposed method achieves high classification accuracy and outperforms other analogous steganalysis approaches.

中文翻译：

---

扩展卷积神经网络作为空间图像隐写分析的特征选择器–混合分类方案

摘要

如今，虽然隐秘术是非法秘密通信的主要手段，但检测隐密术内容（尤其是隐密图像）的需求变得越来越强制。由于多媒体内容可以很容易地在Internet上传播，并且利用了不同领域中更复杂的隐写术算法，即空间，变换，因此识别隐身图像的任务变得非常困难。早期的隐写分析方法对隐密图像进行统计攻击，而较新的方法则使用深度学习技术。后者主要利用卷积神经网络并显示出令人鼓舞的结果。在本文中，我们提出了一种新颖的方法来识别针对两种嵌入速率的隐写图像，该图像是从两种不同的隐写算法S-UNIWARD（空间通用WAvelet相对失真）和WOW（小波获得权重）得出的。所提出的方法最初利用膨胀的卷积神经网络作为特征提取器，然后提取的特征向量训练随机森林分类器。更具体地，证明了在隐写分析中，扩张的卷积神经网络可能是出色的特征提取器，而传统的softmax层可以用另一个机器学习分类器代替。进行了广泛的实验，并将所提出的模型与空间图像隐写分析中使用的最新卷积神经网络以及其他特征提取方法进行了比较。