A New Steganography Algorithm Using Hybrid Fuzzy Neural Networks

Available online at www.sciencedirect.com

ScienceDirect Procedia Technology 24 (2016) 1566 – 1574

International Conference on Emerging Trends in Engineering, Science and Technology (ICETEST - 2015)

A New Steganography Algorithm Using Hybrid Fuzzy Neural Networks Saleema.A,Dr.T.Amarunnishad Mtech,TKM College of Engineering,Kollam,Kerala,India Retd. Principal,TKM College of Engineering,Kollam,Kerala,India

Abstract In recent years, image steganography has been one of the emerging research areas. As the field of information technology is advancing, the need of information security is increasing day by day. Steganography is a widely used communication method in today’s scenario which involves sending secret information in appropriate carriers. Since it have an interesting property of concealing the message as well as the existence of the message, steganography is on its evolutionary path to unearth new platforms. As the field of steganalysis is growing exponentially, the need of developing strong steganographic algorithms is also growing. Since the use of steganography is spreading across various fields, the goal of increasing the embedding capacity, security and image quality is being major concerns.We propose a new image steganographic method which is based on random selection of pixels for secret data embedding and post processing the stego-image using Hybrid Fuzzy Neural Networks. The pixels where secret data is to be embedded is selected randomly using a pseudo random key. In the selected pixels the last 2 or 3 bits are used for hiding. The resultant degradation in the quality of stego-image is handled by an efficient pixel adjustment process with the use of fuzzy neural networks..The experimental results reveal that this method can achieve an embedding capacity of 3 bits per byte with excellent stego-image quality and high imperceptibility. © 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license © 2016 The Authors.Published by Elsevier Ltd. (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-reviewunder under responsibility of the organizing committee of ICETEST – 2015. Peer-review responsibility of the organizing committee of ICETEST – 2015 Keywords:Digital Image Steganography; Imperceptibility: Fuzzy Neural Networks

1. Introduction Steganography deals with the art of hiding information with an interesting property of hiding the mere existence of the secret information. What makes steganography more preferable than cryptography is its extra layer of security provided by the non detect ability of the presence of secret information. Cryptography is the practice of scrambling a

2212-0173 © 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Peer-review under responsibility of the organizing committee of ICETEST – 2015 doi:10.1016/j.protcy.2016.05.139

A. Saleema and T. Amarunnishad / Procedia Technology 24 (2016) 1566 – 1574

1567

message to an obscured form to prevent others from understanding it while steganography is the study of obscuring the message so that it cannot be seen [1]. Steganography requires two files – the cover/carrier file and the secret file. Various multimedia carriers like audio, video, text, image etc can act as a cover media to carry the secret information. Also the secret can be of any type, which in most cases is converted into a bit stream. The resultant file after embedding can also be called as a stego-file. In image steganography the cover file will be an image and the secret may be plain text, cipher text (or another image). The important issues to be considered in steganography are the payload capacity, security/robustness, imperceptibility and computational complexity of insertion and extraction processes [2,3]. 2. Recent Trends in Image Steganography In recent years , researchers have tried to incorporate various intelligent technologies into the field of steganography which resulted in better achievements. Some of them used soft computing tools which turned to be a very big leap in the history of steganography. Such algorithms could result in robust, low cost, optimal and adaptive solutions in data concealing problems. Fuzzy Logic (FL), Rough Sets (RS), Adaptive Neural Networks (ANN), Genetic Algorithms (GA) Support Vector Machine (SVM), Ant Colony, and Practical Swarm Optimizer (PSO) etc. are the various components of soft computing and each one offers specific attributes. In [4], Chang et al presented a novel, reversible steganographic method, which can reconstruct an original image effectively after extracting the embedded secret data.Their method could achieve greater embedding capacity and reversibility. In [5], Nameer proposed a hiding method which used adaptive image filtering and adaptive image segmentation with bits replacement on the appropriate pixels. The main objective was to hide a large amount of any type of information through bitmap image using maximum number of bits per byte at each pixel. Further in [6], the authors introduced a high secure neural based algorithm with an additional layer of security. They introduced a neural network based learning system which helped in increasing the security against statistical and visual attacks. Later in [7], they came up with an intelligent technique using hybrid adaptive neural networks with modified adaptive genetic algorithm. We propose a steganographic algorithm which produces stego images of high quality in statistical as well as visual means with high embedding capacity and low computational complexity. Apart from the conventional models we introduce an additional layer of security by adding a post processing stage for processing the stego image. In this phase of post processing we use a hybrid fuzzy neural network for improving the quality of the stego image.

Fig. 1. Layout of the Proposed System.

1568

A. Saleema and T. Amarunnishad / Procedia Technology 24 (2016) 1566 – 1574

3. Proposed System 3.1. Data Embedding Phase The embedding procedure consists of some preprocessing of the cover image as well as the secret image, secret key generation and data concealing. The selection of the cover image is done based on some inferences from some previous researches on steganography. It should be done wisely so that the stego image formed after concealing will preserve quality [8]. Using a random key, we select some pixels or blocks from the cover image. The secret image is compressed and encrypted before embedding Compression is done to reduce the amount of data to be hidden while encryption is to improve the security. Even if an attacker compromised the primary concern of steganography our data shouldn’t be leaked. Here compression of the secret image is done using SPIHT algorithm[9], which is a powerful wavelet based compression method. Encryption is done using some simple bit operations like AND and OR. We can improve the security by encrypting using strong algorithms like AES or DES. The secret image is then converted to a bit stream and we call it as the secret data. Now in the selected pixels, we hide the secret data using LSB substitution method [10] . If the amount of data is less, it is sufficient to substitute the last 2 bits. Otherwise lest significant 3 bits of the selected pixels are substituted with the secret data to form the stego image. 3.2. Quality Enhancement Phase The stego image formed may not be good in quality which can be improved by post processing using some intelligent hybrid system. This phase is implemented to reduce the chances of statistical detection and some other varieties of image manipulation attacks. The technique used consists of a Hybrid Fuzzy Artificial Neural Network with Back propagation learning algorithm [11,12]. The algorithm used is as follows: Step 1: Implement the data embedding algorithm to conceal the secret message and produce the stego image Step 2: Feature extraction (Finding out some statistical and visual measures for each stego and cover images. Step 3: Create a buffer containing the bits which are not used by the steganographic algorithm. These are bits where secret data is not hidden and we call it as free bits. Step 4: Use a hybrid fuzzy ANN with back propagation with the input layer containing the free bits buffer, statistical and visual measures of the stego image and the output layer containing the new free bits buffer and new statistical and visual measures for the updated stego image. Step 5: Compare the output of the network with those of the cover image. If matching is satisfied, the new modified free bit output in the buffer is used to form a new stego image by assembling it with the other bits where the secret is hidden.Otherwise we will readjust the connection weights using back propagation and repeat from step 4. The operational structure of the network used is depicted in figure 2 [7]. From the stego image formed by concealing, two buffers are extracted, which are the free bits buffer and the buffer with bits holding the secret message. Also the statistical and visual features of the stego images and cover images are extracted and stored. The free bits buffer and the two features of the stego image are given as input to the network. The obtained results are compared with the features of the cover image. If matching is satisfied we can assemble the modified free bits in the free bits buffer with the secret bits to form a new stego image. Otherwise back propagate and adjust the weights until a matching is satisfied. The network we used here is composed of two parts; the fuzzy processing part and a conventional back propagation network [13]. In the fuzzy processing part, the inputs are first processed using a membership function and the processed data will be inputted directly to the back propagation network for further processing. The output data will be compared with the expected output and reversely adjusted based on the mean squared error to specify network connection weights. Feature extraction and fuzzy preprocessing Feature extraction is a special form of dimensionality reduction where the input data (stego image) is transformed into a set of features called feature extraction. We extracted the chi square probability as the statistical feature and

1569

A. Saleema and T. Amarunnishad / Procedia Technology 24 (2016) 1566 – 1574

the Euclidian norm as the visual feature. The same features are extracted from the cover image. and used as the target. The features are extracted using the following methods. 1)chi square probability : We used a method proposed by Wetsman and Pitsfield [14] to extract this feature. This is based on the expected and observed frequencies of the image pixels [15]. 2)Euclidian norm The input features are subjected to fuzzy processing. Based on some analysis it is found out that bell shaped function can well approximate our input space. Suppose that x and y is the input features. As the values of the parameters change, the bell-shaped functions vary accordingly, thus exhibiting various forms of membership functions.

Fig. 2. operational structure of the network.

Suppose that

x0 , y0 are

the inputs. Let

A1, A2

and

B1, B2 be

the linguistic values for each input variable. The

linguistic states can be expressed as in the equations 1 and 2 [16].

ª 1 § u  a ·2 º i1 ¸ » Ai (u) exp« ¨¨ «¬ 2 © bi1 ¸¹ »¼ ª 1 § u  a ·2 º i2 ¸ » Bi (u) exp« ¨¨ 2 b «¬ © i 2 ¸¹ »¼ Where

^ai1, ai2 , bi1, bi2` is the parameter set.

Hidden layer processing

(1)

(2)

1570

A. Saleema and T. Amarunnishad / Procedia Technology 24 (2016) 1566 – 1574

Here we used sugeno reasoning method. Each layer in node which is called the rule nodes calculates the firing strength of the associated rule. We used the logical and operator to get the firing strength as the output of that layer. Let D1 and D 2 are the firing strengths. It can be calculated using the equations 3 and 4.

D1 A1x0 š B1 y0 D2 A2 x0 š B2  y0

(3) (4)

Defuzzification layer Normalization of the firing level of node is performed and then the output neuron is calculated as the product of the normalized firing level as in equations 5 and 6.

E1

E2

D1 D1  D2 D2 D1 D2

(5)

(6)

Then the product of the normalized firing level and the individual rule outputs are calculated using the equations 7 and 8. E1 * z1 E1(a1 * x0  b1 * y0 ) (7)

E2 * z2 E2 (a2 * x0  b2 * y0 )

(8)

Error function The error function can be given by equation 9

E



1 u y o 2



2

(9)

Where y is the desired output and o is the computed output by the hybrid neural net. Example The illustration in figure 3 clearly shows the concept of the quality enhancement procedure using the hybrid FANN. First consider a 3x3 neighborhood in the cover image. The shaded boxes in the figure represent the secret bits to be hidden. The larger boxes show the pixel values of the cover image in binary in which last 2 bits are used for hiding. The edges represent the pixel value difference. Our objective is to reduce the pixel value difference between the neighboring pixels so that it improves the imperceptibility ratio of the stego image.Step1 shows the embedding of the secret data to the 3x3 neighborhood of the cover. Step 2 shows the values after embedding the data to the cover. We can see that some edge values are significantly increased. Step 3 shows adjusting the free bits appropriately which may reduce the edge values. Step 4 shows the improved values after using FANN.The improved stego image will be of high imperceptibility and resilient to statistical and visual attacks by the use of the hybrid system. 3.3. Extraction Phase Extraction is just the reverse process of embedding. Using the pseudo random key generator at the receiver side, the blocks/pixels where the data concealed is find out and the secret data is extracted in the form of bit stream. After converting it to image form, decryption and decompression is performed to get the secret image.

A. Saleema and T. Amarunnishad / Procedia Technology 24 (2016) 1566 – 1574

4. Experimental Results In this section, the performance analysis of the proposed method using Hybrid Fuzzy ANN is performed. Also its comparison with the method using ANN is done. A set of 200 cover images and 50 secret images is used for testing the proposed system. The image dataset consists of smooth as well as complex images. Among this 175 images are used for training and the remaining are used for testing. The data concealing is done and stego images are produced. The same training data set is given to both the networks. All the experiments were performed on a personal computer with specification (2.8 GHz Intel core, 4GB memory and 1TB hard disk with window 8 operating system) and Matlab version R2010a. Figure 4 shows the results of data embedding process. It appears that the resulting stego image is not satisfactory in statistical as well as visual quality. Based on some calculations and observations, we can conclude that the stego image formed is statistically as well as visually very poor.Enhancement of the stego image is performed using Hybrid Fuzzy ANN and the performance was analyzed. The performance was measured by the following means.

Fig. 3. Illustration of the enhancement method

1571

1572

A. Saleema and T. Amarunnishad / Procedia Technology 24 (2016) 1566 – 1574

Fig. 4. Data embedding result. (a) the input image and (b) secret image (c) the result after embedding.

4.1. Testing Image Quality The image quality after concealing was measured using PSNR (equation 10) and structural similarity metric[17] (equation 11).

PSNR 10u log10

max2 MSE

(10)

where max is the maximum pixel value, and MSE represents the average of mean square errors for RGB colors. Table 1 shows the PSNR in dB of some of the stego images without quality enhancement and with enhancement using hybrid Fuzzy ANN. It is found out that the stego image formed by the proposed intelligent technique shows a higher PSNR than that of the stego image without using the hybrid system. The SSIM algorithm (Wang et al., 2004)is used to measure the similarity between two identical images [16]. SSIM(Ic , IS )

where images and

((2PIc PIs  (224 1) u 0.01)2 )((2V Ic Is  (224 1) u 0.03)2 ) PIc2  PIs2  ((224 1) u0.01)2 V Ic2 V Is2  ((224 1) u0.03)2





(11)



PIc PIs are the mean of cover and stego image respectively, V Ic Is is covariance of cover and stego V Ic , V Is are the variance of a cover and a stego images respectively. Table 2 shows the SSIM value of

some of the stego images without quality enhancement and with enhancement using Hybrid Fuzzy ANN.

Stego Images

Secret Images

PSNR without HFANN(average of 3 color planes)

PSNR using HFANN(average of 3 color planes)

Lena

Peppers

36

39.47

F16

Lena

34.02

38.09

Baboon

Tiffany

37.67

40.41

Bardowl

Lena

35.02

38.12

Anhinga

F16

33

36.54

Table 1. Comparison of PSNR in dB

A. Saleema and T. Amarunnishad / Procedia Technology 24 (2016) 1566 – 1574

Stego Images

Secret Images

SSIM before using HFANN

SSIM after using HFANN

Lena

Peppers

0.72

0.8

F16

Lena

0.77

0.87

Baboon

Tiffany

0.78

0.84

Bardowl

Lena

0.8

0.87

Anhinga

F16

0.79

0.89

1573

Table 2. Comparison of SSIM value.

4.1. Amount of Data Hiding The proposed algorithm achieves an embedding capacity of 3 bits per byte with high stego image quality. Many color images are used to test the embedding capacity with different amounts of payload. It is found out that image like bardowl and baboon could conceal largest amount of data with high imperceptible ratio while images like tulips and F16 could conceal lesser amounts. For example, when experiments are performed with Baboon and Bardowl by hiding secret color images of size 256X256 using the proposed steganography scheme based on the intelligent technique, the PSNR value obtained are 39.9 and 42.01 respectively. But for images like F16, when secret color images of size 256X256 are used, thePSNR obtained was 36.07 which is less satisfactory. From these observations, we have inferred that images with smoother regions show lesser embedding capacity than those with complex regions. Also visual quality is better for complex images than smoother images. This is due to the fact that our human vision is sensitive to slight changes in smoother areas while it can tolerate more severe changes in the edge regions. 4.2. Avoiding Statistical and Visual attack The algorithm proposed by Wetsman and Pittsfield [14] is used for statistical stegananlysis. Here we check how the expected and the observed frequencies of stego image pixels are uniformed. It is found out that after concealing the probability value is almost around zero and in most cases the probability value of stego image resembles that of the cover image.As a visual measure Euclidian norm is calculated, which gives the distance between the cover and stego images. It is observed that the distance when the proposed algorithm is used is lesser than when it is not used. 4.2. Speed of Training We have used 200 images for our learning purpose. The same training dataset is given to both networks (simple ANN and hybrid Fuzzy ANN). It is observed that our proposed system learns quicker than the one using ANN. Also the HFANN in our proposed method finds out a match very quickly. Below is the time comparison when a particular input is given.Time Taken Comparison (in seconds) is done.Time taken on Enhancement with NN = 56.501781 seconds while time taken on Enhancement with HFANN =1.434236 secondsFrom the experimental results, it is evident that our proposed method outperforms the existing methods in satisfying the basic steganographic requirements with more focus on image quality enhancement. 5. Conclusion In this work a new steganographic method has been developed for data embedding in color images which better satisfies the trade off between the basic steganographic requirements. Apart from satisfying the same embedding capacity of today’s leading steganographic algorithms, this method concentrates on providing high quality stego image which is highly resistant to statistical as well as visual attacks. Our proposed work makes use of the

1574

A. Saleema and T. Amarunnishad / Procedia Technology 24 (2016) 1566 – 1574

conventional LSB technique for data embedding and introduces an intelligent hybrid technique for stego image enhancement. The experimental results of the proposed intelligent technique are compared with the method using ANN. The results show that our proposed scheme outperforms the existing schemes when standard measurements are taken. The main advantage of our proposed system is that the learning using Hybrid Fuzzy ANN requires a very less time compared to the ordinary neural networks. Also our experimental results clearly show that the stego images formed using our algorithm is not detectable by the state of the art steganalytic methods. References [1] Mansi S. Subhedara, Vijay H. Mankarb, “Current status and key issues in image steganography: A survey”, ScienceDirect, c o m p u t e r s c i e n c e r e v i ew 1 3 – 1 4 ( 2 0 1 4 ) 9 5 – 1 1 3. [2] Petticolas, F.A.P Anderson,R.J.Kuhn,M.G.1999-“Information Hiding-A survey”, Proceedings of the IEEE, Special Issue on Identification and Protection of Multimedia content 87, 1062-1078. [3] Rafael C.Gonzalez,Richard E.Woods,Digital Image Processing,Third Edition. [4] Chang, C., Lin, C., Fan, Y., 2008. "Lossless data hiding for color images based on block truncation coding". Pattern Recognition 41, 23472357, Elsevier. [5] EL-Emam, N., 2007. "Hiding a large amount of data with high security using steganography algorithm". Journal of Computer Science (4),223232, doi:10.3844/jcssp.2007.223.232. [6] El-Emam, N., 2008." Embedded a large amount of information using high secure neural based steganography algorithm". International Journal of Information and Communication Engineering 4 (2), 95106. [7] Nameer N. El-Emam, Rasheed Abdul Shaheed AL-Zubidy, “New steganography algorithm to conceal a large amount of secret message using hybrid adaptive neural networks with modified adaptive genetic algorithm,Elsevier, The Journal of Systems and Software 86 (2013) 1465 1481. [8] Sajedi, H., Jamzad, M., 2010. Boosted steganography scheme with cover image preprocessing. Expert Systems with Applications 37 (12), 7703–7710, Elsevier, doi:10.1016/j.eswa.2010.04.071. [9] Amir Said and William .A. Pearlman , “ A New Fast and Efficient image Codec Based on Set Partition in Hierarchical trees. IEEE transactions on Circuits and Systems for video technology,volume 6 No:3 June 1996. [10] C.K. Chan, L.M. Cheng, Hiding data in images by simple LSB substitution, PatternRecognit. 37 (3) (2004) 469–474. [11] S Sivanandam , S Sumathi, Introduction to neural networks using matlab 6.0. [12] Jinsa K. , and Gunavathi K., "Lung cancer classi_cation using neural networks for CT images" ,Elsevier, computer methods and programs in biomedicine , pp. 202 209, 2014. Shami Pokhrel, and Zhang Jie1, “The Use of Fuzzy [13] Liu Li, Huo Liqing, Lu Hongru, Zhang Feng, Zheng Chongxun, BackPropagation Neural Networks for the Early Diagnosis of Hypoxic Ischemic Encephalopathy in Newborns”, Hindawi Publishing Corporation, Journal of Biomedicine and Biotechnology Volume 2011, Article ID 349490, 5 pages , doi:10.1155/2011/349490. [14] Westfeld, A., Pfitzmann, A., 2000. Attacks on Steganographic Systems. 3rd International Workshop, Lecture Notes in Computer Science, vol. 1768. Springer-Verlag, Berlin, Heidelberg, New York. [15] Christy.A.Stanley, “Pairs of Values and Chi Square Attack”, Department of Mathematics, Iowa State University, May 1, 2005Fachinger, J., 2006. Behavior of HTR Fuel Elements in Aquatic Phases of Repository Host Rock Formations. Nuclear Engineering & Design 236, 54. [16] Robert Fuller, "Neuro-Fuzzy Methods,Tutorial pdf", Lisbon, August 31 and September 1, 2001. [17] Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E., 2004. Image quality assessment from error measurement to structural similarity. IEEE Transaction on Image Processing 13, 600–612, doi:1057-7149/04$20.00.