A CNN and BILSTM based Image Steganalysis

Download Full-Text PDF Cite this Publication

Text Only Version

A CNN and BILSTM based Image Steganalysis

Amina S

CSE Department Valanchery, Malappuram

Mubeena A K

CSE Department Valanchery, Malappuram

Shireen M T

CSE Department Valanchery, Malappuram

Abstract:- In the recent years, deep learning is widely used for steganalysis purpose.in the existing CNN based techniques are effective for the steganalysis than the previous ensemble classifier methods. In this paper, we address this issue by investigating the chance of abusing an organization for steganalyzing pictures of fluctuating sizes without retraining its parameters. On the suspicion that regular picture commotion is comparable between various picture sub regions, we propose a finish to-end, deep learning, novel answer for recognizing steganography pictures from typical pictures that gives fulfilling execution.in this project, we first take the input image, then considers the relation among the noise of various sub regions, and then classification process is done for the sub regions. A CNN bilstm is a hybrid bidirectional LSTM (long short-term memory) and CNN architecture. The CNN component is used to induce the character-level features. Our algorithm adopts a Siamese, CNN-based architecture and a bilstm, which consists of two symmetrical subnets with shared parameters, and contains three phases: pre-processing, feature extraction, and fusion or classification. We used datasets consisting stego images with variable sizes and their corresponding cover images from boss base 1.01 and ALASKA


Keywords: Image steganography, Convolutional neural network., Bidirectional LSTM, Siamese CNN.


Steganography is a technique for concealing privileged information, by installing it into a sound, video, picture or text record. It is one of the strategies utilized to shield mysterious or delicate information from malignant assaults. On the off chance that two clients traded media records over the web, it would be harder to decide if these documents contain covered up messages, then if they were conveying utilizing cryptography.

The picture chose for this design is known as the cover and the picture got after steganography is known as the stego. Figure 1 shows the original image, but there is some data hiding in the figure 2. But it cannot be identified by using normal methods. Only the sender and receiver were aware of it. Thus, it

Fig 1: original image

Fig 2,3: corresponding stego image

can only be decrypted by the recipient owner. Various image steganographic methods like UNIWARD, WOW, etc are used.

In the course of recent years, deep learning has arisen as a promising system giving cutting edge execution for picture steganalysis. Nonetheless, on the grounds that resizing the picture diminishes the sign to-commotion proportion between the cover picture and the steganographic signal, to the most awesome aspect our insight there are at present a couple of best-in-class approaches that can steganalysis discretionary size pictures.

In this paper, we propose a start to finish, exceptionally discriminative neural organization for separating steganographic pictures from unique pictures, one which gives a seriously fulfilling execution in the assessment of pictures of different sizes.

Our network receives a Siamese, CNN-based design, which is used to catch connections between picture sub-regions by utilizing two administrative signals at the same time; those are at that point utilized as the reason for characterization. This will be clarified in more detail in Section III. We can misuse these pieces of information with the understanding that steganographic ally inserted changes in various sub-regions – themselves produced by fulfilling steganography strategies – are not the equivalent. These interior contrasts can give some substance autonomous data for our organization. For the motivations behind preparing and assessing our proposed model, just as creating a preparation dataset, we utilize four picture steganography strategies: S-UNIWARD [2]. The cover pictures were sourced from two information bases: BOSS base 1.01. To encourage future picture steganalysis research, subtleties of both the code and the created dataset will be delivered. The procedure of our Siamese structure could control further investigation into making existing organizations handle heterogeneous datasets all the more viably.

The remainder of the paper is coordinated as follows: in Section II, we momentarily audit related work tending to a few significant regards; in Section III, we present our CNN-based model, and examine its hidden destinations; we make determinations on the discoveries introduced in this paper.


    Our point is to introduce a novel, viable answer for steg-analysing pictures of self-assertive size. Appropriately, in this part we survey applicable, common picture steganography draws near what's more, existing CNN-based steganalysis strategies.

    • Most common steganographic approaches

    The expectation hidden image steganography draws near is to impact an undetectable change to cover a mystery message in a cover picture. Regardless of whether an outsider finds the stego, doubts concerning the information covered up inside it are improbable on the grounds that it resembles an ordinary picture. The most widely recognized picture steganography draws near can be comprehensively classified into three classes: naive steganography [3], [4], versatile steganography [2][6], and deep learning-based installing.

    Naive steganography strategies are moderately straightforward and are utilized generally across the Internet for amusement purposes. These are likewise the strategies that make the most effectively discernible ancient rarities. For example, the Least-Significant-Bit (LSB) strategy [3],

    1. inserts a mysterious message into the cover picture by changing the estimations of pixels without estimating the contortion. Accordingly, they can be handily assaulted basically by utilizing the earlier measurable information on cover pictures.

      Adaptive steganography is at present the most down to earth technique. It not just improves security by installing mysterious messages into more finished regions of the cover pictures [2][6], be that as it may, it likewise utilizes effective steganographic codes, for example Syndrome Trellis Codes (STCs) [10], to limit the aggregate effect of the inserted changes. For example, Pevný et al. [5] characterized a weighted contrast of highlight vectors utilized in steganalysis in their investigation of contortion in their Highly Imperceptible steGO (HUGO) strategy. Holub and Fridrich [5] developed a model of individual pixel costs in Wavelet Obtained Weights (WOW) by breaking down the adjustment in the yield of directional high-pass channels delivered by evolving one pixel. After a year, they streamlined the added substance distortion based on directional residuals got utilizing a channel bank in their

      UNIversal WAvelet Relative Distortion (S-UNIWARD) device [2]. Li et al. [3] proposed an expense work for the HIgh-pass, Low-pass and Low-pass (HILL) strategy by utilizing a high-pass channel to find the less unsurprising parts, and utilizing two low-pass channels to make the minimal effort esteems more grouped.

      Deep learning-based implanting is a developing examination field. There are four significant groups of profound learning-based installing (1) Via union: this kind of technique generates pictures and afterward shrouds the message in them (2) Generating a likelihood guide of changes: in models definite by Tang et a. [9] (ASDL-GAN) and Yang et al. [8] (UT-6HPF-GAN), the generator network creates a modification map from the cover picture, which makes it possible to misdirect the discriminant network. (3) Via tricking amazing CNN-based steganalyzers [8], [4]: for instance, (4) Via 3-player game [2], [9]: for example, Zhu et al. [10] as of late developed joint train encoder and decoder (HiDDeN) organizations – given an info message and cover picture, the encoder produces an outwardly undefined encoded picture, while the decoder can likewise disentangle the first message from it. Until this point, the advancement of profound learning-based installing strategies is still at an undeveloped stage, yet there are indications of potential.

      • Deep learning based steganalysis

    Using deep learning and neural network image steganalysis can be made with higher accuracy. CNN-based steganalysis strategies dispense with the requirement for hand-created highlights and consequently separate more broad highlights from information using backpropagation.

    Qian et al. [11] proposed an altered CNN model (which they called GNCNN), an effective worldview made out of three sections: preprocessing, feature extraction and classification layers. Boroumand et al. [12] proposed SRNet, an alternate worldview that arbitrarily introduces all channels and that

    Figure 3: System design

    Figure 4: different layers in the convolutional neural network

    separates satisfactory commotion lingering through different, unpooled convolutional blocks. SRNet is perhaps the best methodologies as of now accessible for high- recognition exactness.

    Albeit these techniques are portrayed by essentially better execution contrasted and regular identifiers, they can't straightforwardly prepare on enormous pictures because of the limits of existing equipment. Because of the identity of feeble steganographic signals [9], resizing or trimming the pictures preceding grouping will bargain the precision of the indicator. Until now, barely any strategies have been created trying to tackle this issue. Hong Zhang [3] proposed Siamese CNN for distinguishing stego and cover images.

    In this work, we propose an alternate steganalysis answer for pictures of discretionary size for CNNs with serious location exactness. We

    Figure 5: Stego images displayed

    exploit the connections between picture sub-locales to recognize whether a picture is a stego.


    Deep learning is a moderately incredible strategy. Nonetheless, when varieties are made to the substance, size or lighting of the picture, or to the shooting gear, the cover pictures may look altogether different. One of the principal difficulties of steganalysis is to create strong administrative signals to conquer these modifications. In this paper, we recognize connections between picture sub-locales when the steganography activity. Our proposed method consists of

    Figure 6: Training graph indicates accuracy and loss

    three sections: pre-processing, feauture extraction and classification.

    Firstly, the input image is vertically divided into two sub areas. Each of these were passed into two subnets which shares same weights, structures and parameters. At the front of each subnet, the preprocessing stage is utilized to create picture commotion residuals (n), which are profoundly identified with the steganographic signal. Then, the element extraction stage is utilized to separate the element vector (f) of each sub-territory clamor leftover. The system architecture is shown in figure 3. The image is subdivided into two sub areas and each sub areas passed to identical subnets for feature extraction. From each sub regions patches are created for much effective classification. And bidirectional LSTM is used to classify the noise residuals from the images. Then by combining statistical features extracted from the two subnets and bilstm a fully connected layer introduced for classification. BiLSTMs successfully increment the measure of data accessible to the organization, improving the setting accessible to the calculation. A BiLSTM layer learns bidirectional long-haul conditions between time steps of time arrangement or succession information. These conditions can be valuable when you need the organization to gain from the total time arrangement at each time step.

    In the pre-processing stage, noise is extracted using the convolutional layers. Features are extracted during the second stage. Here we obtained a 128-dimensional vector as an output. Fusion or classification is the final stage the two subnets output are concatenated and calculated and obtained a 512-dimensional vector.

    Figure 8: weighted AUC garaph

    subnets output are concatenated and calculated and obtained a 512-dimensional vector.A softmax layer is used as the fully connected layer. Figure 4 represents the neural network created during the training process. The confusion matrix is used to calculate the accuracy and the AUC (area under curve) is obtained. From these values it is clear that our proposed system is much more efficient.


    Training and testing: we get the cover images from BOSSBASE, and convert them into stego images using UNIWARD (universal wavelet relative distribution) and WOW (wavelet obtained weights) method. for the conversion of certain cover images into stego, payload is added to the image vectors. The payload is added to 128 x

    128 sized images and is converted to 512 x 512 sized vectors. After the training process of the convolutional layers, the graph is plotted which indicates the loss and accuracy of our method. Figure 6 shows the loss and accuracy of the proposed network when training and validating 256 sized images from BOSS base. S-UNIWARD method is used and the technique is implemented in the MATLAB. WOW method is also used for the comparison with BILSTM. It is indicated in figure 8. Then by concatenating the features obtained from Siamese architecture and bilstm , a 512 feature size picture obtained. This is fed to a fully connected layer and classification layer for further classification. In this step checked whether the image vectors are similar or not- ie., to check whether the image is stego or not. A graph is plotted by combining the two stego methods with the proposed method. From the graph it is clear that the accuracy

    Figure 7: Roc curve of our proposed network when testing images for S-UNIWARD, WOW and LSTM

    of the system is increased. That is by using bilstm along with Siamese CNN the network is much more effective. The auc value obtained is about 0.948. Thus, our proposed system easily distinguishes the stego images from datasets of variable sizes. Average value of the accuracy is calculated to optimize the result.


Image steganalysis is the method used for decoding the hidden data or information from an image. A tale steganalysis technique for the picture of subjective size is proposed, which is exclusively founded on a Siamese convolutional neural organization (CNN) that abuses the connections between picture sub- regions. Our proposed Siamese, deep learning system is a novel answer for recognizing stego pictures from cover pictures. Not only Siamese CNN but also BILSTM is used for image classification. Thus, the system is more efficient. Our proposed network is well-generalized and robust among the previous methods. One of the major applications of this method is it can be used efficiently in the forensic area or police departments to control various crimes among terrorists.


  1. Weike You , Hong Zhang and Xianfeng Zhao A Siamese CNN for Image Steganalysis Ieee Transactions On Information Forensics And Security, Vol. 16, 2021august 15,2020

  2. V. Holub and J. Fridrich, Digital image steganography using universal distortion, in Proc. 1st ACM Workshop Inf. Hiding Multimedia Secur. IH&MMSec, 2013, pp. 1719.

  3. R. G. van Schyndel, A. Z. Tirkel, and C F. Osborne, A digital watermark, in Proc. 1st Int. Conf. Image Process., Nov. 1994, pp. 8690.

  4. R. Cogranne, C. Zitzmann, L. Fillatre, F. Retraint, I. Nikiforov, and

    P. Cornu, A cover image model for reliable steganalysis, in Proc. 13th Int. Conf. Inf. Hiding, Prague, Czech Republic, May 2011, pp. 178192.

  5. T. Pevný, T. Filler, and P. Bas, Using high-dimensional image models to perform highly undetectable steganography, in Proc. 12th Int. Conf. Inf. Hiding (IH), Calgary, AB, Canada, Jun. 2010, pp. 2830.

  6. T. Filler, J. Judas, and J. Fridrich, Minimizing additive distortion in steganography using syndrome-trellis codes, IEEE Trans. Inf. Forensics Security, vol. 6, no. 3, pp. 920935, Sep. 2011.

  7. J. Yang, D. Ruan, J. Huang, X. Kang, and Y.-Q. Shi, An embedding cost learning framework using GAN, IEEE Trans. Inf. Forensics Security, vol. 15, pp. 839851, 2020.

  8. S. Bernard, T. Pevný, P. Bas, and J. Klein, Exploiting adversarial embeddings for better steganography, in Proc. ACM Workshop Inf. Hiding Multimedia Secur., Jul. 2019, pp. 216221.

  9. J. Hayes and G. Danezis, Generating steganographic images via adversarial training, in Proc. Annu. Conf. Neural Inf. Process. Syst. (NIPS), Long Beach, CA, USA, Dec. 2017, pp. 19541963.

  10. V. Sedighi, R. Cogranne, and J. Fridrich, Content-adaptive steganography by minimizing statistical detectability, IEEE

    Trans. Inf. Forensics Security, vol. 11, no. 2, pp. 221234, Feb. 2016.

  11. Y. Qian, J. Dong, W. Wang, and T. Tan, Learning and transferring representations for image steganalysis using convolutional neural network, in Proc. IEEE Int. Conf. Image Process. (ICIP), Sep. 2016, pp. 2528.

  12. M. Boroumand, M. Chen, and J. Fridrich, Deep residual network for steganalysis of digital images, IEEE Trans. Inf. Forensics Security, vol. 14, no. 5, pp. 11811193, May 2019

Leave a Reply

Your email address will not be published. Required fields are marked *