An Enhanced Digital Image Watermarking Scheme for Medical Images using Neural Network, DWT and RSA

Published Online April 2016 in MECS DOI: 10.5815/ijmecs.2016.04.06

Digital form products like video, audio, text etc. are being transmitted in everyday of our lives [1, 2]. In the field of digital multimedia communication, some techniques has helped a lot for storing, editing and accessing of these products [3]. But, security during the transferring communication is very crucial. As lack of security leads to the loss in property rights. So to solve this problem, digital watermarking is very good [4]. Fig. 1 shows the progression of watermark embedding. Digital watermarking is the process of embedding an image with secret data for the communication. The embedded image can only be extracted by person who has authentication.

Fig.1. Watermark embedding process

There are two types of image watermarking, one is blind and the other one is non-blind. In blind image watermarking, there is no requirement of original content and it is also called as public image watermarking. In non-blind image watermarking, there is requirement of original content and it is also known as private image watermarking.

• A.    Medical Images

The amount of digital images in the internet has been increasing day by day. The need of secure diagnosis in the medical field is very necessary. As these days transmission of images has become common, so security is needed at high rate.

Medical images are used to view tissues, organs or body parts of a person. These images are used for disease diagnosis and treatment of the disease. Medical images play a major role in the classification of diseases. These are also helpful in monitoring disease activity. Several machines are accessible today that can generate images of internal organs of a person. There are different types of imaging methods such as x-ray, ultrasound, MRI scans, CT scans, Nuclear medicine scans etc [16]. Electromagnetic waves are also called x-rays. X-rays are the common form of medical imaging. X-rays are passed through the body of a person to produce images of internal structure of body. X-rays are commonly used for checking swallowed objects, blood vessels, broken bones etc. Ultrasound is also known as medical sonography. Sound waves are used in this method to check internal organs. It is used in pregnancy or to check organs in abdomen and pelvis etc. MRI stands for Magnetic Resonance Imaging. Radio waves and magnetic field are used in this method so that organs and tissues of the body can viewed in much more detail. It is useful to check spinal injuries, abnormal tissues, blood vessels etc. CT scans are similar to X-rays but in this technique, multiple X-rays are projected from different angles to obtain a 3-D view of different organs. Medical images must be taken care of to confirm the identity as all diagnose depends on the obtained medical image. In order to fulfill the security and convenience issues of the patients watermarking is used in this field.

• B.    Noise and Distortions

K = imnoise(O,'localvar',B) (1)

It adds zero-mean, Gaussian white noise of local variance B to the image O. B is an array of the same size as O. Rotation attack is the most common attack that has been found these days in an image. The obtained image is rotated through several angles like 90 degree, 180 degree and 360 degree [8].

In this paper, a new method is proposed in which encrypted watermark is embedded in the medical image by using DWT and Neural Network approach. Section II describes previous work done in this field. Section III describes the techniques used in proposed work. Section IV describes the stimulation model. Results and comparisons are presented in section V. Finally, section VI describes conclusion and future scope [17].

• II.    Related Works

Digital watermarking is an outstanding field of research. A lot of researchers have proposed large number of algorithms in this field. The main purpose of all these algorithms is to provide imperceptibility and robustness against different attacks. Several types of watermarking techniques found in the academic literature are presented in this section.

Saxena, P. et al. [12] proposed a method for watermarking based on the usage of the SVD-DWT. The main advantage of this method is that it is robust to various types of attacks. In proposed work the watermark is embedded in high frequency band by SVD. Kaur, R. et al. [14] presented the new watermarking algorithm based on the usage of DWT-SVD techniques. In proposed work 4 sub-bands will be obtained using DWT and then watermarking will be done using Singular Value Decomposition method. From experiments it has been concluded that proposed method has better efficiency. Sridevi, T. et al. [15] proposed the watermarking method in which instead of taking an original image a reference image has been taken. Then with the help of fuzzy logic the image is watermarked in the reference image. Here use of fuzzy logic is done to embed the image and the use of SVD is done to enhance the embedding process after the application of DWT for the decomposition of an image into 4 levels. In the end results are performed using NCC and PSNR metrics.

Koley, S. et al. [16] proposed a new type of information hiding ability in biomedical images by using the combination of cryptography and digital watermarking to attain the improvement in confidential and authenticated data storage and secured transmission. They use patient’s name and doctor’s name as patient’s information and it is encrypted using cryptography and embedded in the patient image by using watermarking. Higher order bit LSB replacement technique is used for embedding the information and RSA is used for encryption. The private keys are also embedded in the cover image to have enhanced security and precise recovery of the concealed information. The outcome of the proposed methodology shows that the concealed information doesn‘t influence the cover image and it can be recovered efficiently even from several noisy images. The strength of the proposed embedding scheme is also supported by numerous image quality matrices. Solanki, N. et al. [17] proposed an algorithm in which work is divided in two main stages; initially image is divided into two parts: Region of Interest i.e. defined as informational piece of image and Region of Non Interest i.e. defined as non-informational piece. They have used Region of Non Interest to conceal the data so that the informational part of the image remains protected. RSA is used for encryption of watermark. A Discrete Wavelet Transform based approach is utilized for embedding the encrypted data in Region of Non Interest part of image for improving data security. The experimental result demonstrates that watermark embedded by the proposed algorithm gives better security and can survive successfully under different attacks.

Anita, et al. [18] presented a review on watermark coding and decoding technique for image security. DWT technique is utilized for coding as DWT have broad range of functionalities. Two images are used, one as a host image and second as a watermark image. Then the image is extracted to retrieve the watermark image. This all is done under DWT technique. They basically evaluate the DWT technique, its improvement, and its challenges and also result are analyzed on MATLAB tool by computing SSIM, PSNR and SSR values of extracted image. Sridhar, S. et al. [19] combine the features of watermarking, image encryption in addition to image steganography to give consistent and protected data transmission .The fundamentals of data hiding and encryption are enlightened. The initial step entails inserting the mandatory watermark on the image at the best possible bit plane. The next step is to utilize RSA hash to encrypt the image. The last step entails attaining a cover image and hiding the encrypted image inside this cover image. A set of metrics will be used for evaluation of the effectiveness of the digital watermarking. The list incorporates Mean Squared Error, Peak Signal to Noise Ratio and Feature Similarity.

Venkatram, N. et al. [20] proposed an algorithm for authentication of medical images. Medical images restrain extremely susceptible information. Watermarking medical images entail cautious amendments protecting the data in the images. RSA algorithm is used for encryption and decryption of patient image. Medical images like MRI, CT and Ultrasound scans of body parts of patients are used as a cover image. 2D DWT (Discrete Wavelet Transform) is used to watermark the encrypted patient image. The medical cover image and watermark image are transformed into wavelet domain and are diverse using two scaling factors alpha as well as beta. At last these watermarked medical images are passed on through the internet in conjunction with the secret key which will be used for decryption. On the receiving end the embedded encrypted watermark is extracted using 2D DWT and decryption key. The robustness of the projected watermarking techniques is tested with various attacks on the watermarked medical images. Peak-Signal-to-Noise ratios (PSNR) and Normalized cross correlation (NCC) coefficients are computed to access the eminence of the watermarked medical images and extracted patient images. The results are evaluated for three types of medical images with one patient image watermarks using single key by using four wavelets (haar, db, symlets, bior) on four diverse levels (1,2,3,4).

• III.    Tecniques Used in Proposed Work

• A.    Discrete Wavelet Transform (DWT)

The discrete wavelet transform is a valuable way designed for signal exploration as well as picture handling, chiefly in multi-resolution description. It can crumble signal into different components in the frequency sphere. It divides an image into 4 sub-bands i.e. HL, HH, LH and LL. One-dimensional discrete wavelet transform (1-D DWT) decomposes an input into two components (the average component and the detail component). The 2-dimensional (2-D) DWT distributes an input picture into four type of sub-bands, single average component (LL) and three detail components (LH, HL, HH) as shown in Fig. 3.

Fig.3. DWT second level decomposition

The utilization of wavelets is done mainly for the denoising of an image or to decompose the image into its constituent forms [11, 12]. Detailed has LL and LH. Image consists of pixels that are arranged in two dimensional matrix, each pixel represents the digital equivalent of image intensity. DWT transforms the spatial domain into frequency domain. Below equations shows the mathematical portion of a discrete signal:

g 1[ n 1] =

M 1

∑ W 1 φ [ j 1 0 , k 1] φ 1 j 0, k [ n 1] + k 1

∞

¹              W 1 ψ [ j 1, k 1] ψ 1 j , k [ n 1]

^{M 1} j 1 = j 1 o k 1

h ere , g 1[ n 1], φ 1 j 1 o , k 1 [ n 1] and ψ 1 j 1, l 1 are discrete functions which are defined in [0, M-1], totally M1 points. For the reason that the sets {ϕ1 j1o,k1 [n1]} k1ƐC and {ψ1 j1,k1 [n1]} (j1,k1)ƐC1 ² ,j1≥k1o , are orthogonal to each other.

• B.    RSA Algorithm

  RSA stands for Rivest, Shamir and Adelman. It turned out 1st for sale in 1978 among the 1st public key cryptographic programs. RSA can be used pertaining to key alternate along with digital signatures. Currently, RSA is generally utilized to encrypt your procedure key useful for secret key encryption (message integrity) or message's hash value (digital signature). It is the most widely used algorithm in the world for cryptography. It can be used for public key cryptography as well as digital signatures. In this, Party A send the message to party B without prior knowledge to party B then using decrypt key B can extract and read the message sent by party A.

Within cryptography, RSA is an algorithm pertaining to public key cryptography. The particular RSA algorithms entail the application of two keys. A public key which is often known by everyone, in addition to enable you to encrypt statement. A private key, acknowledged simply because of the receiver, in addition utilizing to decrypt communication. To produce a great RSA public/private key match, there are some principle measures:

Step 1: Opt for two prime numbers, p1 as well as q1. Commencing these numbers we can compute the modulus, n=p1*q1.

Step 2: Choice a third number, e, which is reasonably prime to (i.e., it doesn’t distribute uniformly into) the product (p₁-1) (q₁-1). The number e may be the public exponent.

Step 3: Compute an integer d through the quotient (ed-1)/ [(p 1 -1) (q 1 -1)]. The number d may be the private exponent.

Step 4: Compute an integer d through the quotient (ed-1)/ [(p 1 -1) (q 1 -1)]. The number d may be the private exponent.

Step 5: The public key may be the number pair (n, e). Although these kinds of valuations are generally widely acknowledged, it truly is computationally infeasible to discover d from n addition to e if p1 in addition to q1 are generally big enough.

To encrypt a communication message, M, while using the public key, build your cipher text, C, while using the equation:

C = Me mod n (3)

The receiver after that decrypts your cipher text while using private key using the equation:

M = Cd mod n (4)

• C.    Neural Network

The Back Propagation neural network is artificial neural network based on error back propagation algorithm. The Back Propagation (BP) neural network model consists of an input layer, more or less hidden layers as well as an output layer. Each connection connecting neurons has a distinctive weighting value. In training the network, the nodes in the BP neural network obtain input information from exterior sources, and then go by to hidden layer which is an interior information processing layer and is answerable for the information conversion, and then the nodes in the output layer supply the required output information. After that, the back-propagation of error is transported by distinct the actual output with wanted output.

BP neural network consists of many neurons that are arranged in a form of three layers: input, hidden and output. The neurons are linked by weights W y In training the network with a given architecture, the back propagation approach, finds a single best set weight values by minimization of suitable error function. In a multi-layer feed forward neural network, the processing elements are arranged in layers and only the elements in adjacent layers are connected. It has a minimum of three layers of elements (i.e., input layer, the middle or hidden layer, and the output layer).

The name “back propagation” (BP) derives from the fact that computations are passed feed forward from the input layer to the output layer, resulting which computed errors are broadcasted back in other direction to change the weights to obtain a better performance. BP algorithm is an extension of the least mean square algorithm that can be used to train multi-layer networks.

Once the network weights and biases are preliminary, the linkage is prepared for training. The preparation process necessitates a group of instances for proper network behavior, such as network inputs p and destination outputs t. For the duration of training the weights as well as biases of the network are iteratively adjusted to minimize the network performance function. The number of hidden layer is always difficult to determine in ANN creation. It is generally agreed that one hidden layer is sufficient for most of purposes. In the present study, only 1 hidden layer will, thus, be used in the BP network for simplicity.

Neural network consists of various no. of neurons and their working is similar to the brain neuron structure. There are various types of neural networks but commonly used neural network is Back Propagation Neural network [14]. The normal Back propagation is the mainly applied to train Multilayer FNN.

• IV.    Stimulation Model

The proposed idea will be implemented in MATLAB which is extensively utilized in all regions of applied mathematics, in education as well as research by universities, and in the industry. The flowchart for proposed work is shown in Fig. 4. The methodology of proposed work is given as:

Fig.4. Flowchart of proposed work

• A.    Training Panel

Initially images are uploaded for Training. After that calculation of image bits has done. Now the lower and upper DWT values are calculated. Load DWT values for training and after that training is done using Neural

Network. Fig. 5 shows the Neural Network training process.

• B.    Testing Panel

In testing panel, initially cover image is uploaded. After that watermark image is uploaded. Now implementation of RSA will be done to encrypt watermark. The encrypted watermark is embedded in cover image.

Fig.5. Neural Network training process

• C.    Extraction Panel

In extraction panel, watermarked image is uploaded and neural network is applied to obtain extracted watermark. In the end parameter evaluation will be done using five metrics i.e. PSNR, MSE, BCR, BER and NCC.

• V.    Results and Discussions

In this section, experimental results of the proposed algorithm are evaluated. The imperceptibility and robustness are evaluated by using different parameters and different attacks are applied to test the robustness of the proposed algorithm.

• A.    Results

  The proposed scheme is implemented in MATLAB 7.10.0 (R2010a). The proposed scheme is tested on six gray scale and six colored medical images with size 512×512. A gray scale watermark of size 64×64 is used. The performance of the proposed scheme is evaluated on the basis of PSNR, MSE, NCC, BER and BCR. These parameters are used to check the quality of the watermarked image. Different attacks are applied on the watermarked image to check the robustness. Table 1 show the gray scale watermarked medical images and watermarked medical images with different attacks. Different parameters are evaluated for watermarked

image and watermarked image with different attacks to check the imperceptibility and robustness of the proposed algorithm. Similarly, table 2 shows the colored watermarked medical images and watermarked medical images with different attacks. Different parameters are evaluated for watermarked image and watermarked image with different attacks to check the imperceptibility and robustness of the proposed algorithm.

Table 1. Experimental results of gray scale medical images

WATERMARKED IMAGE			PARAMETERS	WATERMARKED IMAGE WITH ATTACKS		PARAMETERS
			MSE =0.2899	** h -^—^ — *V*i A •	Q p	MSE =0.5687
	"^ »ta—», ^----"t**^/.		PSNR =54.5382			PSNR =51.7634
			BER =0.3298			BER =0.2663
			BCR =4.6129			BCR =4.9221
			NCC =0.9985			NCC =0.9941
	Hi		MSE =0.4676		4? Z (ZI s 4?	MSE =0.6953
			PSNR =52.9416			PSNR =50.4185
			BER =0.5589			BER =0.5812
			BCR =5.0266			BCR =5.1838
			NCC =0.9955 MSE =0.4219			NCC =0.9964 MSE =0.5198
	w^^l		PSNR =53.9430			PSNR =52.8832
			BER =0.3827			BER =0.4453
			BCR =5.0484			BCR =5.3625
			NCC =0.9942			NCC =0.9940
			MSE =0.5064	■ ^VsWlB' ■	(Z) Z	MSE =0.6368
			PSNR =51.8941			PSNR =50.5347
			BER =0.6570			BER =0.6609
			BCR =4.8782			BCR =5.3688
			NCC =0.9901			NCC =0.9929
			MSE =0.5301	. *?W V / ;           - ^. .:	0	MSE =0.6120
			PSNR =52.6875			PSNR =51.3476
			BER =0.6736			BER =0.4624
			BCR =5.0112			BCR =5.0881
			NCC =0.9969			NCC =0.9948
	@^ M		MSE =0.5124		&? у	MSE =0.5899
			PSNR =52.1182			PSNR =51.4627
			BER =0.5172			BER =0.5358
			BCR =5.0486			BCR =5.0314
			NCC =0.9914			NCC =0.9919

Table 2. Experimental results of colored medical images

WATERMARKED IMAGE

PARAMETERS

MSE =0.2885

PSNR =54.5670

BER =0.3298

BCR =4.6115

NCC =0.9915

MSE =0.5828

PSNR =51.6083

BER =0.5593

BCR =5.1417

NCC =0.9985

MSE =0.4642

PSNR =53.1430

BER =0.3829

BCR =5.0907

NCC =0.9973

WATERMARKED IMAGE WITH ATTACKS

PARAMETERS

MSE =0.6393

PSNR =51.0726

BER =0.2666

BCR =4.9927

NCC =0.9928

MSE =0.7084

PSNR =50.8047

BER =0.5842

BCR =5.2331

NCC =0.9958

MSE =0.5807

PSNR =50.9606

BER =0.5779

BCR =4.9292

NCC =0.9988

MSE =0.3047

PSNR =54.7452

BER =0.5290

BCR =4.7648

NCC =0.9955

MSE =0.5874

PSNR =51.4436

BER =0.5940

BCR =5.2307

NCC =0.9956

MSE =0.5831

PSNR =51.4919

BER =0.4292

BCR =5.4859

NCC =0.9926

MSE =0.5742

PSNR =52.0150

BER =0.6738

BCR =5.2506

NCC =0.9945

MSE =0.6667

PSNR =50.6501

BER =0.4588

BCR =5.1949

NCC =0.9937

MSE =0.7561

PSNR =49.9271

BER =0.6155

BCR =5.3961

NCC =0.9937

Without Noise

—■- With

Noise

Images

Fig.6. Graph of MSE values

Without

Noise

—•- With Noise

Images

Fig.7. Graph of PSNR values

Values of BER for different colored medical images are shown in Fig. 8. Imperceptibility of the proposed algorithm is evaluated on the basis of parameters of watermarked images and robustness of the proposed algorithm is evaluated on the basis of different attacks applied on these watermarked images. From the above graph, it can be observed that the values of BER for all images are between 0.2 and 0.7 and lesser value of BER gives enhanced results. This shows that the proposed scheme has good imperceptibility and robustness.

Values of BCR for different colored medical images are shown in Fig. 9. Different attacks are applied on the watermarked images to check the robustness of the proposed algorithm. This robustness ensures that the data can be transmitted securely despite various attacks they might be subjected to. The embedded watermark cannot be tampered easily by using different attacks. Imperceptibility is also important so that informational part of the image does not get distorted. From the above graph, it can be observed that the values of BCR for watermarked images and watermarked images with different attacks are between 4.6 and 5.6 and larger value of BCR gives improved results. This shows that the proposed scheme has good imperceptibility and robustness.

-■- With

Noise

Images

Without

Noise

Fig.10. Graph of NCC values

Values of NCC for different colored medical images are shown in Fig. 10. From the above graph, it can be observed that the values of NCC for all images are close to 1 and higher value of NCC gives enhanced results. This shows that the proposed scheme has good robustness.

• B.    Comparison of Results

The comparison of existing scheme and proposed scheme is shown in table 3 [20]. The values of NCC for the proposed algorithm are higher than the existing method and higher value of NCC gives better results. These results demonstrate that the proposed scheme has more robustness.

Table 3. Comparison of results

Type of Attack	NCC (Existing Method)	NCC (Proposed Method)
Without Noise	0.9835	0.9985
Salt & Pepper Noise	0.6038	0.9959
Gaussian Noise	0.6061	0.9971
Rotation Attack	0.9653	0.9936

Fig. 11 shows the graph of NCC for existing and proposed method. It can be observed that the values of NCC for proposed method are higher than the existing method and values of NCC which are close to 1 gives better results. Therefore, the proposed algorithm gives enhanced results than the existing method.

Fig.11. Comparison of results

The comparison of existing scheme and proposed scheme is shown in table 4 [21]. Three medical images are used for testing and values of PSNR for all three images are higher than the existing method. These results demonstrate that the proposed scheme has more imperceptibility.

Table 4. Comparison of results

Type of Attack	PSNR (Existing Method)	PSNR (Proposed Method)
Image 1	42.48	56.96
Image 2	48.46	50.31
Image 3	42.52	50.85

Fig. 12 shows the graph of PSNR for existing and proposed method. It can be observed that the values of PSNR for proposed method are higher than the existing method. Therefore, proposed method gives enhanced results than the existing method because higher value of PSNR gives better results.

■PSNR (Exisitng)

■PSNR (Proposed)

Image 1

Image 2

Image 3

Fig.12. Comparison of results

The comparison of existing scheme and proposed scheme is shown in table 5 [21]. The values of PSNR for proposed method are higher than the existing method. These results demonstrate that the proposed scheme has more robustness and imperceptibility.

Table 5. Comparison of results

Type of Attack	PSNR (Existing Method)	PSNR (Proposed Method)
Without Noise	42.52	50.85
Salt & Pepper Noise	31.94	50.56
Gaussian Noise	30.03	50.34
Rotation Attack	35.87	50.32

Fig. 13 shows the graph of PSNR for existing and proposed method. It can be observed that the values of PSNR for proposed method gives enhanced results than the existing method because higher value of PSNR gives better results.

Fig.13. Comparison of results

• VI.    Conclusion and Future Scope

In this paper, DWT is used to get low and high pass filter values. Then, training of images is done using Neural Network. RSA algorithm has been used for encryption of watermark to make it more secure. After that, embedding and extraction of watermark is done using Neural Network. Results are evaluated using five parameters such as PSNR, MSE, BCR, BER and NCC.

In future, hybridization of different techniques can be used to enhance results.

Acknowledgment

The authors wish thanks to all research experts of the Information Technology Department of CEC and Computer Science Department of CGC, Landran, Mohali, Punjab, India.