International Journal of Social Ecology and Sustainable Development

Volume 13 • Issue 1

Image-to-Image Steganography
Using Encoder-Decoder Network
Vijay Kumar, National Institute of Technology, Hamirpur, India*
Ashish Choudhary, National Institute of Technology, Hamirpur, India
Harsh Vardhan, National Institute of Technology, Hamirpur, India

ABSTRACT

In this paper, a convolution neural network is utilized for image-to-image steganography using encoder
decoder architecture. A new loss function is designed to improve the invisibility of payload image.
The encoder-decoder architecture is used for the image-to-image steganography. The developed
architecture is evaluated on the well-known image dataset and compared with the recently developed
models. The proposed model was able to withstand the stegoanalyzer attack and better visual quality
of stego image. It is able to achieve great imperceptibility.

Keywords
Deep Learning, Generative adversarial networks, Image Steganography, Steganalysis, Stego-image

1. INTRODUCTION

Due to the development in technology, the transmission of data from one place to another is easy
and fast. Whereas, the breach of information can be done through the advance tools and techniques.
Sometimes, the leaked information may cause the severe losses. Information hiding techniques are
used to resolve this problem (Kumar and Kumar, 2010). These techniques are able to conceal the
important data in a way so that the intruder is unable to reveal the secret data. These are widely used
in business and army for secret data communication. These techniques are broadly categorized into
three classes namely, watermarking, cryptography, and steganography (Girdhar and Kumar, 2018). In
watermarking, the watermark is added to the data for their authenticity. The watermark can be text,
image, and audio. Watermark can be visible or invisible according to the applicability of watermarking
in a specific area. However, this technique reveals the presence of watermark and easily modified by
the intruders (Kaur et al., 2020). The second well-known technique is cryptography. Cryptography
technique encrypts the secret message itself. This technique scrambles the secret message so that the
intruders are unable to reveal the important information from the scrambled message (Al-Ataby and
Al-Naima, 2010). The third technique is steganography. Steganography conceals the data into another
media for the security purpose. This technique uses encoder to encode secret data into cover data.
The decoder is used to decode the encoded message for secret message extraction. According to the
nature of data, steganography techniques are broadly categorized into three classes namely, image,
audio, and video (Kumar and Kumar, 2019). The main focus of this paper is image steganography
due to the ease to implement in various domains.

DOI: 10.4018/IJSESD.312181 *Corresponding Author


Copyright © 2022, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.


1
International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

Image steganography can be implemented in two different domains. These are spatial and
frequency domains (Kumar et al., 2020). The former one modify the pixel values of an image through
the computational techniques. Least-Significant Bit (LSB) technique is the well-known example of
spatial domain technique (Ker, 2005). LSB manipulates the lowest-order bits of pixels in the given
images to conceal the secret message. It methodically amends the statistical distribution of pixels of
the image. The latter one transform the pixel values of the image. The transformed coefficients were
processed to encode the secret message (Kumar and Kumar, 2017). The well-known frequency domain
techniques are discrete cosine transform (DCT) and discrete wavelet transform (DWT). However,
these techniques are not sufficient to handle the secret message in this modern era. To develop the
efficient steganography technique, the concepts of modern technologies have been incorporated.
Recently, the deep learning techniques are widely used in steganography for strengthen the concealed
secret data. The well-known deep learning technique is the convolutional neural network (CNN). Baluja
(2017) utilized CNN for hiding the image into another cover image. An auto-encoder is used for image
compression during the hiding process. Rehman et al. (2017) developed an encoder-decoder network to
conceal the grey images into other images. In this paper, an encoder-decoder methodology is designed
for the image steganography. This model can easily concealed the secret image inside a cover image
and also extracted it with low distortion in quality. The main contributions of this paper are:

1. An encoder decoder methodology is designed for image steganography with automatic feature selection.
2. An extensive hyper-parameter tuning is done, which majorly lags in the previous research works.
3. The designed methodology is evaluated on the publicly available datasets and tested through the
well-known evaluation measures.

The remaining structure of this paper is as follows. Section 2 presents the related work done in
the direction of image steganography. The proposed steganography method is mentioned in Section
3. The experimental results and discussion are mentioned in Section 4. The future research directions
are presented in Section 5 followed by the conclusion in Section 6.

2. RELATED WORK

A lot of research has been done in the field of image steganography using deep learning techniques.
The emergence of generative adversarial networks (GAN) can be applied in the image steganography.
Due to this, researchers utilized GAN in this field.
Rehman et al. (2017) developed an encoder-decoder network for image steganography. A new
loss function was designed to train the developed network. The proposed network was tested on five
datasets. This model provided better results than the other techniques. Setiadi (2019) utilized the most
significant bit pixels of cover to hide the secret image. This method attained better PSNR value as
compared to the existing techniques. The embedding capacity of proposed method was also increased
by a large number of pixels. However, the computational complexity of the proposed method is high.
Swain (2019) proposed two steganography techniques based on the difference between pixel value
and quotient value. The neighbouring pixels were computed according to central pixel. The proposed
technique attained better value of PSNR than the existing techniques. However, the performance of
this method can be further improved by compression and encryption techniques.
Das et al. (2021) used the deep neural networks to encode multiple images into a single cover
image. However, this method has high loss value. New loss functions have to be designed to improve
the performance of this method. Kumar et al. (2020) used conventional neural network (CNN) to
design the steganography technique. Adam optimizer was used in CNN. It was able to provide the
better visual perception for stego image. Ray et al. (2021) hybridized deep learning and edge detection
techniques for image steganography. This method embedded the payload data in the edge area of

2
 International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

an image and less data was embedded in non-edge areas. This method provided better embedding
capacity as compared to the other techniques. Hamid et al. (2021) used SqueezeNet for embedding
the secret data into the cover image. They used the concepts of S-UNIWARD method. This method
is independent from the features of dataset. The classification accuracy obtained from this method
was better than the other techniques. Fu et al. (2020) proposed a model named HIGAN, which was
based on the encoder-decoder network. The secret image was extracted through decoder. This model
attained less colour distortion and high security.
Shang et al. (2020) used adversarial concepts to improve the image steganography. This method
was to able withstand the attacks from steganalyzer tools. It was able to extract the secret image
with less distortion. Zhang et al. (2019) employed a GAN to hide the grey images into Y channel of
cover image. This method minimized the divergence between stego and cover images. A new loss
function was designed to generate better stego image. This method suffers from the security problem.
Baluja (2017) developed a CNN-based on the organization of encoder-decoder network. The encoder
network concealed the secret image into the cover image. Whereas, the decoder network extracted
the secret image from the encoded image. This method distorted the colour quality of stego images.
Volkhonskiy et al. (2016) developed a new model for steganography using deep convolutional GAN
(DCGAN). This model was able to provide realistic stego image. Hayes and Danezis (2017) utilized
the concepts of neural networks as a component in the traditional algorithms. They used deep neural
networks to model the data-hiding pipeline. These networks have significantly improved the efficiency
in terms of maintaining the secrecy and quality of the encoded messages.
Al-Ataby and Al-Naima (2010) proposed a modified high-capacity image steganography technique
that depends upon wavelet transform with acceptable levels of imperceptibility and distortion in the
cover image. Zhu et al. (2018) developed an end-to-end framework for image steganography. Three
CNNs were used for encoder, decoder, and adversary networks. This framework was able to generate
the better stego image as compared to other techniques. Shi et al. (2017) utilized the concepts of
adversary-based secret image embedding and detecting game.

3. PROPOSED MODEL

The proposed model for steganography is motivated from Rehman et al. (2017). Two modifications
have been done in the existing technique. First, the payload image is taken from the same dataset
instead of other source. Second, the leaky ReLU is used in the proposed model. The encoder-decoder
methodology is used in the proposed model. By using this technique, it eliminates the need of feature
selection. The description of the proposed model is mentioned in the succeeding subsections.

3.1 Model Pipeline

A pair of encoder-decoder CNNs is trained to generate the new hybrid image from both cover and
payload images. Both cover and payload images are taken from same dataset, i.e., CIFAR-100 dataset.
The payload image can be recovered from the generated hybrid image. Figure 1 illustrates the cover
(or source) and payload images, which are taken from CIFAR-100 dataset.
Figure 2 shows the pipeline of proposed model. The encoder module extracts precise features from
the cover image to conceal the details from the payload images. This module takes cover and payload
images to generate the hybrid image. The produced image is visually indistinguishable to the cover
image and also have the contents of payload image. The decoder module is used to differentiate the
hidden features from the generated hybrid image. This module is used to extract the payload image
from the generated hybrid image. The visual contents of extracted image is comparable to the cover
image generated from the encoder module.

3
International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

Figure 1. Source and Payload Image from CIFAR -100 dataset

Figure 2. Encoder decoder pipeline

3.2 Model Architecture

We also perform a lot of experiments with other design choices; however, this architecture comes out
as the best choice after comparing the results so we finalise this architecture .The architecture of our
model is mentioned in Figure 3. A brief description of our encoder/decoder architecture is as follows:
The encoder network utilizes the aggregation of two parallel branches. These are cover and payload
branches. Each of these branches receive different inputs. One receives the cover image and other
receives payload image. The cover branch is made up of 11 separate layers. Among them, eight are
made from combination of Conv2D layer and ReLU layer. The remaining two have the combination
of Conv2D layer and Leaky ReLU layer. The last layer has a single Conv2D layer. The main aim of
Cover Branch Filter is to segment the given input image into the hierarchical representation of features.
Thereafter, they merged the features of cover image into payload image. The payload branch is made
up of eight separate layers that are made from the combination of Conv2D layer and ReLU layer. The
purpose of this arrangement is to extract the both high and low-level features from the given image.
The architecture of decoder network is the aggregation of seven layers. The first four layers are
made up of Conv2D layers followed by a ReLU. The next two layers are made up of a combination
of Conv2D and Leaky ReLU. The last layers of decoder network is formed through a single Conv2D
layer. This module is responsible for extracting the concealed payload from the generated hybrid image.

4
 International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

Figure 3. Architecture of proposed model (a) Encoder network (b) Decoder

3.2.1 Loss Function

In the previous work done by Rehman et al. (2017), mean squared error (MSE) was used as the
loss function in the model training and the optimizer used for training the model was Adam. In the
proposed model, both loss function and optimizer are modified to enhance the performance of model.
After the lot of calibration, mean absolute error (MAE) is used as loss function. As MAE-based loss
function provides high accuracy and more robust towards the outliers. For optimizer, RMSprop is
used as an optimizer in place of Adam. As RMSprop uses a moving average of squared gradients to
normalize the gradient itself. This will be helpful to balance the step size. Leaky RELU is used in
the proposed model.

5
International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

3.3 Working of Model

The working of the proposed model is given below:

1. Dataset is decomposed into two sub-datasets by using the split size of 0.70 for training and 0.30
for testing purpose.
2. Payload and cover images are randomly selected from the corresponding dataset.
3. The transformers are used to normalize the data.
4. Set the control parameters of proposed model. The learning rate of RMSprop optimizer is set to
0.001. 150 epochs are used to train the model. The batch sizes of training and testing modules
are set to 128 and 64, respectively. Both the encoder and decoder are stacked in different models.
5. Train the proposed model on the training dataset.
6. Test the proposed model on testing dataset.

4. EXPERIMENTAL RESULTS

The performance of the proposed model is evaluated on a diverse set of publicly available datasets namely Tiny
ImageNet (Abai and Rajmalwar, 2020), CIFAR-100 (Krizhevsky and Hinton, 2009), and F-MNIST (Xiao et
al., 2017). The well-known performance measures are used to validate the performance of proposed model.

4.1 Datasets Used

CIFAR-100 dataset have 60000 colour images with each size of 32x32. These images are categorized into
hundred classes and each class have 600 images. 500 images per class are used for training purpose and 100
images per class are used for testing purpose. Figure 4 shows the sample images from CIFAR-100 dataset.

Figure 4. Some sample images from CIFAR -100 dataset

The second dataset is F-MNIST. It consists of 60,000 images for training purpose and 10,000
images for testing purpose. Each image has size of 28x28. These images are categorized into ten
different classes. Figure 5 illustrates the sample images from F-MNIST.
Tiny ImageNet dataset consists of 100,000 images and each image have size of 64x64. These
images are grouped into 200 classes. Each class contains 500 images for training purpose. 100 images
are used for validation and testing purpose. Due to limited computation power available, we have
used 10,000 images for experimentation. The some sample images are depicted in Figure 6. Table 1
shows the detail descriptions of the above-mentioned datasets.

6
 International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

Figure 5. Some sample images from F-MNIST dataset

Figure 6. Some sample images from Tiny ImageNet dataset

Table 1. Detail description different datasets used

Name of Dataset Size of Images Total No. of Images

Tiny ImageNet 64 x 64 1,00,000
Fashion MNIST 28 x 28 70,000
CIFAR - 100 32 x 32 60,000

4.2 Evaluation Metrics

The robustness of proposed steganography model is evaluated through distortion and secrecy. For
this, Structural Similarity (SSIM) index and Peak Signal to Noise Ratio (PSNR) are used to evaluate
the visual quality of the image. The mathematical formulation of PSNR is given below (Kumar and
Kumar, 2017):

 Max 2 
 I 
PSNR = 10log   (1)
 MSE 

7
International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

where MaxI denotes the maximum possible pixel value of the given image. MSE denotes the mean
square error and is defined as (Wang et al., 2004):

m n
1 2
MSE =
mn ∑∑ I (i, j ) − K (i, j ) (2)
i =1 j =1

where m and n represent the dimension of an image. I and K denote the original image and reconstructed
image, respectively.

4.3 Performance Evaluation

Figure 7 shows the visual analysis of the proposed model on Tiny ImageNet dataset. It is observed from
figure that the visual quality of stegao image is similar to the source image. The extracted image is quite
similar to the payload image. The stego image is not provide any idea that the payload is embedded in it.

Figure 7. Results of hiding payload images into source image on Tiny ImageNet dataset

4.3.1 Quantitative Analysis

We used images from CIFAR -100 dataset as cover and payload image. The image size is set to 32x32.
Initially, all the image in this dataset is coloured image but to make it a payload image we convert the
image into grey image so that it can be become a desirable payload. For F-MNIST dataset, the image size
is set to 28 x 28 pixels. Due to limited computation power, 10000 images are taken from Tiny ImageNet.
These tiny coloured images have the size of 64 x 64 pixels. Table 2 depicts performance comparison in
terms of PSNR and SSIM on different datasets. For F-MNIST dataset, PSNR values for encoder and
decoder obtained from the proposed model are 38.22 and 37.19, respectively. For Tiny dataset, PSNR
values for encoder and decoder obtained from the proposed model are 37.23 and 36.78, respectively. It
shows the better embedding of the data in the cover image and undetectable by the naked eye.

8
 International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

Table 2. Comparison of PSNR and SSIM values on different dataset obtained from the proposed model

4.3.2 Performance Comparison

The proposed model is compared with CNN-based encoder decoder model (CNN-ED) over CIFAR-100
and Tiny ImageNet datasets. Figure 8 shows the comparative analysis of these models in terms of
PSNR. The results reveal that the proposed model attained better PSNR values than the CNN-ED
for both datasets. The analysis of these models is done through the SSIM over CIFAR-100 and Tiny
ImageNet datasets (see Figure 9). The SSIM values obtained from the proposed model are better
than CNN-ED. It is observed from figures that the proposed model outperforms the CNN-ED in
terms of SSIM and PSNR.

Figure 8. PSNR obtained from the proposed model and CNN-ED on (a) CIFAR-100, (b) Tiny ImageNet datasets

Figure 9. SSIM index obtained from the proposed model and CNN-ED on (a) CIFAR-100, (b) Tiny ImageNet datasets

9
International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

4.3.3 Effect of Loss Function

Table 3 shows the effect of loss function and optimizer on the performance of the proposed model.
The proposed model is tested over the combination of ReLU+MSE and Leaky ReLU+MAE. For this,
CIFAR-100 dataset is used. It is observed from table that the combination of Leaky ReLU and MAE
perform better than the combination of ReLU and MSE in terms of PSNR and SSIM. The PSNR values
obtained from the Leaky ReLU+MAE are 38.31 and 36.52 for encoder and decoder, respectively. The SSIM
values obtained from the Leaky ReLU+MAE are 0.986 and 0.972 for encoder and decoder, respectively.

Table 3. Effect of loss function on the performance of proposed algorithm over CIFAR–100

5. FUTURE DIRECTION

In comparison to the earlier methods, the encoder-decoder network is used to either augment or replace a small
portion of secret image in the image hiding system. The work presented in this paper can be improved in near
future. The adversarial networks can be incorporated to enhance the security. The embedding capacity can be
improved by using the computational techniques. The visual inspection can be further improved by using the
latest techniques. This can be achieved by the joining adversarial network in the encoder-decoder architecture.
This will help in making this system more secure and can easily able to bypass any steganalysis software.

6. CONCLUSION

In this paper, the image steganography model is proposed to conceal gray secret image into a color
image with the same size excellently. The proposed model utilized the concept of encode-decoder
network. The loss function has been designed by incorporating the mean absolute error. Leaky ReLU
has been added in the developed model to eliminate vanishing gradient problems. The proposed
model was evaluated on three different image datasets. This model provide better results than the
other technique in terms of PSNR and SSIM.
The proposed model can be further improved by decreasing the loss in decoder network. The image
steganography can be extended with generative adversarial network that will tweak the model and also
boost it security several times.

CONFLICT OF INTEREST

The authors of this publication declare there is no conflict of interest.

FUNDING AGENCY

This research received no specific grant from any funding agency in the public, commercial, or not-
for-profit sectors.

10
 International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

REFERENCES

Abai, Z., & Rajmalwar, N. (2020) DenseNet models for tiny ImageNet classification. ArXiv:1904.10429
Al-Ataby, A., & Al-Naima, F. (2010). A modified high capacity image steganography technique based on wavelet
transform. The International Arab Journal of Information Technology, 7(4), 358–364.
Baluja, S. (2017) Hiding images in plain sight: Deep steganography. In: Proceedings of Advances in Neural
Information Processing Systems, 30, (pp. 2069–2079).
Das, A., Wahi, J. S., Anand, M., & Rana, Y. (2021) Multi-Image steganography using deep neural networks.
ArXiv:2101.00350
Fu, Z., Wang, F., & Cheng, X. (2020). The secure steganography for hiding images via GAN. EURASIP Journal
on Image and Video Processing, 46.
Girdhar, A., & Kumar, V. (2018). A comprehensive survey of 3D image steganography techniques. IET Image
Processing, 12(1), 1–10. doi:10.1049/iet-ipr.2017.0162
Hamid, N., Sumait, B. S., Bakri, B. I., & Al-Qershi, O. (2021). Enhacing visual quality of spatial image
steganography using SqueezeNet deep learning network. Multimedia Tools and Applications, 80(28-29),
36093–36109. doi:10.1007/s11042-021-11315-y
Hayes, J., & Danezis, G. (2017) Generating steganographic images via adversarial training. In: proceedings of
Neural Information Processing Systems, (pp. 1-10).
Kaur, M., Kumar, V., & Singh, D. (2020). An efficient image steganography method using multi-objective
differential evolution. Digital Media Steganography.
Ker, A. D. (2005). Steganalysis of LSB matching in grayscale images. IEEE Signal Processing Letters, 12(6),
441–444. doi:10.1109/LSP.2005.847889
Krizhevsky, A., & Hinton, G. (2009). Learning multiple layers of features from tiny images. Technical Report.
University of Toronto.
Kumar, V., & Kumar, D. (2010) Performance evaluation of dwt based image steganography. In: Proceedings of
IEEE International on Advance Computing Conference, (pp. 223-228). Patiala. doi:10.1109/IADCC.2010.5423005
Kumar, V., & Kumar, D. (2017). A modified DWT-based image steganography technique. Multimedia Tools
and Applications, 77(11), 13279–13308. doi:10.1007/s11042-017-4947-8
Kumar, V., & Kumar, D. (2019). Performance evaluation of modified color image steganography using discrete
wavelet transform. Journal of Intelligent Systems, 28(5), 749–758. doi:10.1515/jisys-2017-0134
Kumar, V., Rao, P., & Choudhary, A. (2020). Image steganography analysis based on deep learning. Review of
Computer Engineering Studies, 7(1), 1–5. doi:10.18280/rces.070101
Ray, B., Mukhopadhyay, S., Hossain, S., Ghosal, S. K., & Sarkar, R. (2021). Image steganography using deep
learning based edge detection. Multimedia Tools and Applications, 80(24), 33475–33503. doi:10.1007/s11042-
021-11177-4
Rehman A.U., Rahim R., Nadeem S., & Hussain S.U. (2017) End-to-End Trained CNN Encoder-Decoder
Networks for Image Steganography. CoRR, abs/1711.07201.
Setiadi D.R.I.M. (2019) Improved payload capacity in LSB image steganography uses dilated hybrid edge
detection. Journal of King Saud University – Computer and Information Sciences.
Shang, Y., Jiang, S., Ye, D., & Huang, J. (2020). Enhancing the security of deep learning steganography via
adversarial examples. Mathematics, 8(9), 1446. doi:10.3390/math8091446
Shi, H., Dong, J., Wang, W., Qian, Y., & Zhang, X. (2017) Ssgan: Secure steganography based on generative
adversarial networks. In Pacific Rim Conference on Multimedia, (pp. 534–544).
Swain, G. (2019). Two new steganography techniques based on quotient value differencing with addition-
subtraction logic and PVD with modulus function. Optik (Stuttgart), 180, 807–823. doi:10.1016/j.ijleo.2018.11.015

11
International Journal of Social Ecology and Sustainable Development
Volume 13 • Issue 1

Volkhonskiy, D., Borisenko, B., & Burnaev, E. (2016) Generative adversarial networks for image steganography.
In: Proceedings of International Conference on Learning Representations, Toulon, France
Wang, Z., Bovik, A. C., Sheikh, H. R., & Simoncelli, E. P. (2004). Image quality assessment: From error visibility
to structural similarity. IEEE Transactions on Image Processing, 13(4), 600–612. doi:10.1109/TIP.2003.819861
PMID:15376593
Xiao, H., Rasul, K., & Vollgraf, R. (2017) Fashion-MNIST: a novel image dataset for benchmarking machine
learning algorithms. ArXiv:1708.07747
Zhang, R., Dong, S., & Liu, J. (2019). Invisible steganography via generative adversarial networks. Multimedia
Tools and Applications, 78(7), 8559–8575. doi:10.1007/s11042-018-6951-z
Zhu J., Kaplan R., Johnson J., & Fei-Fei L. (2018) HiDDen: Hiding data with deep networks. CoRR,abs/1807.09937.

Harsh Vardhan is currently pursuing a Btech from NIT Hamirpur. His research interest is Image Steganography,
Deep Learning, and Machine Learning.

12