**Sanghoon Kang 1, Hanhoon Park 1,\* and Jong-Il Park <sup>2</sup>**


Received: 11 September 2019; Accepted: 23 October 2019; Published: 26 October 2019

**Abstract:** This study proposes a convolutional neural network (CNN)-based steganalytic method that allows ternary classification to simultaneously identify WOW and UNIWARD, which are representative adaptive image steganographic algorithms. WOW and UNIWARD have very similar message embedding methods in terms of measuring and minimizing the degree of distortion of images caused by message embedding. This similarity between WOW and UNIWARD makes it difficult to distinguish between both algorithms even in a CNN-based classifier. Our experiments particularly show that WOW and UNIWARD cannot be distinguished by simply combining binary CNN-based classifiers learned to separately identify both algorithms. Therefore, to identify and classify WOW and UNIWARD, WOW and UNIWARD must be learned at the same time using a single CNN-based classifier designed for ternary classification. This study proposes a method for ternary classification that learns and classifies cover, WOW stego, and UNIWARD stego images using a single CNN-based classifier. A CNN structure and a preprocessing filter are also proposed to effectively classify/identify WOW and UNIWARD. Experiments using BOSSBase 1.01 database images confirmed that the proposed method could make a ternary classification with an accuracy of approximately 72%.

**Keywords:** image steganalysis; WOW; UNIWARD; ternary classification; convolutional neural network (CNN)

## **1. Introduction**

Interest in information security technologies, such as image steganography/steganalysis, has significantly grown because of the universalization of digital multimedia and communication. Image steganography is a technique in which a secret message is embedded into an image, called *cover image*, and the message-embedded image, called *stego image*, is transmitted through a public channel without gaining the attention of a third party, thereby implementing covert communication. The image steganalysis is the reverse process of image steganography, which aims to determine whether or not the image to be tested contains a secret message and then finds out the hidden message.

The performance of image steganographic methods depends on two conflicting parameters: embedding capacity, which represents how many messages we can hide, and the image quality after embedding, which is closely related to message concealment. Therefore, most image steganographic methods have achieved a high embedding capacity at the expense of low image quality after embedding, and vice versa.

Early image steganographic methods include the least significant bit (LSB) substitution method [1], which replaces the least significant bits of image pixels by secret messages, and the pixel value differencing (PVD) methods [2–4] that determine the amount of secret messages to be embedded in

proportion to the difference between adjacent pixels. These early image steganographic methods sequentially embed secret messages into all pixels of an image, although they have been recently extended to embed messages in randomly selected pixels using pseudo-random generators for secure message hiding [5,6].

Sequentially embedding secret messages into all pixels of an image is well known to change the statistical characteristics of the image. In Figure 1, the solid line refers to a probability density function (PDF) of the differences between two adjacent pixels on a cover image. The dot lines refer to the different PDFs on the stego images created by different image steganographic methods. The PDF of the LSB stego image is significantly different from that of the cover image in the section where the differences are small. This statistical difference is easily detected by statistical attacks, such as the RS analysis in [7]. Thus, image steganographic methods have come to consider more how not to be detected by steganalytic attacks than how many messages to embed. To avoid statistical attacks, image steganographic methods began to consider where the message would be embedded. Methods such as HUGO [8], WOW [9], and UNIWARD [10] tried to embed a message into only pixels with a small distortion, mainly on image edges, by analyzing the distortion caused by embedding a message into each pixel. For example, HUGO measured the embedding distortion by reverse-engineering the processes of the subtractive pixel adjacency matrix (SPAM) [11], a steganalytic method that calculated a co-occurrence matrix for the differences of the adjacent pixels in eight directions of vertical, horizontal, and diagonal to analyze the statistical changes in the pixel values caused by the message embedding. HUGO could reduce the probability of being detected by the SPAM by 1/7.

**Figure 1.** Probability density functions of the differences between the adjacent pixels on a cover image and its stego images.

The performance of image steganalysis in detecting image steganography has greatly improved with the development of image steganography to more covertly and skillfully hide a message. Image staganalytic methods generally try to extract traces of image steganography in the image by using high-pass filters (HPF) and identify images to which image steganography has been applied through classification. Early steganalytic methods extracted image features using manually designed HPFs (those features are called handcrafted features hereafter) and detected image steganography using classifiers based on machine learning algorithms, such as support vector machines (SVM) [12] and random forest [13]. A representative method using handcrafted features is the spatial rich model (SRM) [14].

With the great success of convolutional neural networks (CNN) in object detection and recognition [15,16], using CNNs for steganalysis has been actively investigated [17–27]. Unlike handcrafted feature-based methods, a CNN can automatically extract and learn the features that are optimal or well suited for identifying steganographic methods. Therefore, CNN-based steganalytic methods have demonstrated a better performance compared to handcrafted feature-based methods.

However, most existing image steganalytic methods, regardless of whether or not CNNs are used, have focused on identifying whether or not a secret message is hidden in an image (i.e., the binary classification between a normal (or cover) image in which any message has not been embedded and a stego image in which a message has been embedded). Discriminating stego images created by different steganographic methods has been less considered; thus, the binary classifiers are not suitable for discriminating these stego images. Discriminating the stego images created by WOW and UNIWARD that embed a message in a similar and skillful manner is very difficult.

The classification of stego images created by different steganographic methods plays an important role in restoring embedded messages beyond judging whether or not a message is embedded. In this study, as the first step to restore messages embedded by steganographic methods, a CNN-based steganalytic method is proposed to classify the stego images created by different steganographic methods. The structure of a ternary classifier is specially designed to distinguish between the stego images created by WOW and UNIWARD and the normal images without messages. Through comparative experiments with the existing binary classifiers, the reason why multiple steganographic methods should be classified in a single ternary classifier, and various methods for improving the performance of the proposed ternary classifier are presented.

Compared to existing image steganalytic methods, the primary contributions of this study are as follows:


This study is an extension of [28] and differs from the previous study in the following respect:


The remainder of this paper is organized as follows: Section 2 briefly reviews the conventional image steganographic and steganalytic methods; Section 3 explains the proposed steganalytic method; Section 4 experimentally evaluates its performance using images from a database available online; and Section 5 presents the conclusions and suggestions for future work.

### **2. Related Work**

### *2.1. WOW and UNIWARD*

WOW and UNIWARD calculate the degree of distortion when a message is embedded in an image, and then embed a small amount of message in regions where the distortion is small. We refer herein to such methods as adaptive steganographic methods. This makes it more difficult to detect hidden messages by embedding messages only in high-frequency regions with relatively little distortion and makes it possible to avoid steganalytic attacks using statistical analysis because the change in the statistical characteristics of the images caused by message embedding is very small (Figure 1).

Adaptive steganographic methods have suggested different approaches for quantifying the image distortion caused by message embedding. The image distortion function for WOW is defined as follows:

$$D(X,Y) = \rho\_{ij}(X,\mathcal{Y}\_{ij})|X\_{ij} - \mathcal{Y}\_{ij}|.\tag{1}$$

Here, *X* and *Y* are a cover and its stego images, respectively, and *ρ* is a function that examines the detectability in all neighboring directions of each pixel using the HPFs in Figure 2. Thus, a message is not embedded if the detectability is high even in one direction. The message is embedded into the pixels for which the detectability is low in all directions.

$$\begin{array}{cccc} \textbf{KB} & \textbf{K}^{(1)} = \begin{pmatrix} -1 & 2 & -1 \\ 2 & -4 & 2 \\ -1 & 2 & -1 \end{pmatrix} \\\\ \text{Solebl} & \textbf{K}^{(1)} = \begin{pmatrix} 1 & 2 & 1 \\ 0 & 0 & 0 \\ -1 & -2 & -1 \end{pmatrix}, \textbf{K}^{(2)} = (\textbf{K}^{(1)})^T \\\\ \text{WDFB-H} & \textbf{h} = \textbf{Haar wavelet decomposition} \\ \textbf{g} & \textbf{H}\textbf{a} \text{ wavelet decomposition} \\ \textbf{h} & \textbf{D} = \textbf{Dubecchies} \textbf{8} \text{ wavelet decomposition} \\ & \begin{array}{c} 0.5 \\ 0 \\ -0.5 \end{array} \\ \textbf{g} & \textbf{=Daubecchies} \textbf{8} \text{ wavelet decomposition} \\ \textbf{g} & \textbf{= Daubecchies} \textbf{8} \text{ wavelet decomposition} \\ \textbf{0.5} & \textbf{1} \oplus \textbf{1} \\ -0.5 & \textbf{1} \end{array}$$

$$\textbf{K}^{(1)} = \textbf{h} \cdot \textbf{g}^T, \textbf{K}^{(2)} = \textbf{g} \cdot \textbf{h}^T, \textbf{K}^{(3)} = \textbf{g} \cdot \textbf{g}^T$$

**Figure 2.** HPFs and wavelet filters used in WOW [9].

For UNIWARD, the residual images were calculated using the wavelet filters in Figure 2. The image distortion function is defined as follows by the sum of the absolute difference between the cover and the stego residual images:

$$D(X,Y) = \sum\_{k=1}^{3} \sum\_{u=1}^{n\_1} \sum\_{v=1}^{n\_2} \frac{|\mathcal{W}\_{uv}^k(X) - \mathcal{W}\_{uv}^k(Y)|}{\sigma + |\mathcal{W}\_{uv}^{(k)}(X)|}. \tag{2}$$

Here, *<sup>W</sup>*(*k*) represents the residual image calculated using the *<sup>k</sup>*th filter; *<sup>n</sup>*<sup>1</sup> and *<sup>n</sup>*<sup>2</sup> are the image width and height, respectively, and *σ* is a constant stabilizing the numerical calculations.

Consequently, WOW and UNIWARD have different image distortion functions, but their approaches to embedding messages are very similar.
