Machine Learning-Based Robust Watermarking Technique for Medical Image Transmitted Over LTE Network

1 Introduction

In the early 1800s, people used naive methods for transferring information and messages from one place to another, but the transfer of these methods depended on the distance and length of the message. For example, special flags, smoke signs, or short messages and other visual signals were used for small-distance communication. However, to send a message over a long distance, using couriers was a more feasible basic choice. With the span of time, a new era of communication has taken place, where a long message (audio, video, or text) can be easily transmitted over a long distance with less latency [12]. Therefore, the demand for mobile Internet grew exponentially. As per the Opera web browser, the number of pages viewed has raised from about 23 billion in January 2010 to 177 billion pages in November 2013 [37].

Over the past years, the significant change in the usage of smartphones, tablets, netbooks, and laptops with wireless as well as broadband connections has added as many tens of millions of users [25]. These examples show the huge demand for higher transfer rate, better availability, and higher speed for mobile Internet connections. The researchers and the telecom industry have been working hard to come up with new ideas for network architecture at low cost as well as high-speed broadband connection for mobile access. The generation of wireless networks 1G, 2G, 2.5G, 3G, and Worldwide Interoperability for Microwave Access (WiMAX) like network infrastructure has been installed for several applications such as text, voice, and data transfer. This infrastructure is being used in several telemedicine applications. However, lower data rates and secure transmission are the main issues.

The third generation of mobile communication standards is announced by the Third-Generation Partnership Project (3GPP) as Long-Term Evolution (LTE). In mobile communication, LTE is the modern technology. It is considered as a 3.9G mobile network and is being standardized by 3GPP [3]. The title LTE is taken from the Evolved Universal Terrestrial Radio Access Network (E-UTRAN) [1, 2]. It is constituted on all IP frameworks, which is not limited by earlier design. In downlink, the data rate is enlarged to 100 Mbps for data transmission using orthogonal frequency division multiplexing (OFDM), whereas the uplink access is upgraded to 50 Mbps by means of the Single Carrier-Frequency Division Multiple Access (SC-FDMA) modulation scheme. The cellular network bandwidth is ascendable from 1.25 to 20 MHz and improved infrastructure is employed using LTE [4, 14]. Different bandwidth allocations and spectrum availability of different network operators to provide deferent services are accommodated by means of this technique. The LTE technique fully uses the given spectral bandwidth for high data rate and voice services than 3G network carriers.

The 3GPP LTE technology is used to create an unfailing communication for telemedicine applications between hospitals and other medical purposes. The data transmission through open-channel network requires much attention for data security, reliability, and integrity. The secure data ought to be protected by unauthorized access, modification, or deletion. All these requirements can be easily fulfilled by means of steganography. This stimulates learning the use of frequently explored clinical images in securing the patient’s essential data. Therefore, digital image watermarking can be so much helpful for such applications.

These recent advancements in new emerging technologies for the transmission of data through wired or wire communication has opened the gate for intruders. Nowadays, data acquisition, transmission, and exchange are become so simple tasks due to the Internet and the availability of printers, fax, and computers. This transmission of digital data through open-channel network creates opportunities for the third parties or intruders to access data without any permission.

In this paper, we have proposed a novel watermarking technique in addition to the transmission of watermarked image over a simulated OFDM-based LTE physical layer. In this watermarking technique, a machine learning-based algorithm is applied to the cover medical image for the classification of ROI and NROI. A small misclassification in the selection of ROI and NROI can cause big trouble in diagnosis. Considering the LTE/3GPP technology for telemedicine applications, the bit error rate (BER) and the peak signal-to-noise ratio (PSNR) of the received signal have been evaluated.

The result achieved in this experiment revealed that a 10⁻⁶ BER value is realizable for a greater value of SNR (i.e. more than 10.4 dB of SNR). The PSNR of the received cover image is more than 35 dB, which is acceptable for clinical applications. The heartening results of the proposed embedding method can have potential applications in telemedicine.

2 LTE: An Overview

In cellular technology, LTE emerges as a new technology toward fast data communication using the existing allocated frequency spectrum. It enhances all the services such as data rate, voice, and video communication for the end users.

The efficient transmission of data and control information between an enhanced base station (eNodeB) and mobile user equipment (UE) is enlarged by means of LTE [46]. Some new advanced technology and services have been added to the cellular communication such as OFDM, SC-FDMA, and multiple input multiple output (MIMO) [16]. In downlink, the data rate is enlarged using OFDM, whereas the uplink access is upgraded by means of the SC-FDMA modulation scheme in LTE [9]. In OFDMA, a quantified number of symbols are permitted to or from several users on a subcarrier-by-subcarrier basis. The three vital components of LTE network are E-UTRAN, UE, and Evolved Packet Core (EPC; Figure 1 and Table 1).

Figure 1:

System Interface.

The EPC connects data packet to the external world via the Internet or by other means of networking. Uu, S1, and SGi are shown as the interfaces between the different parts of the system.

Intercarrier interference (ICI) and intersymbol interference (ISI) have no effects in OFDM. Higher data bit streams are divided into small parallel bit streams in OFDM and transmitted over a large number of parallel narrow-band subcarriers in spite of a single wide-band carrier [15, 20]. The orthogonality property of OFDM removes any sign of interference among the data streams. In addition, it is robust against narrow-band interference and multipath problem.

The OFDM-based modulation scheme has some demerits. It is sensitive to frequency errors and phase noise, as subcarriers are closely spaced. It is delicate to Doppler shift and causes interference between the subcarriers (ICI) [36]. Another demerit of OFDM is the high peak-to-average signals, which introduces SC-FDMA in uplink, recompensing the disadvantage of OFDM. The functional commonalities of SC-FDMA and OFDMA are shown in Figure 2.

Figure 2:

Functional Commonality of SC-FDMA and OFDMA.

1. Constellation mapper: Depending on channel conditions (BPSK, QPSK, or 16QAM), it changes the arriving bit stream into single carrier symbols.

2. Serial/parallel converter: As an input to the FFT engine, it sets up time domain SC symbols into blocks.

3. M-point DFT: It converts time domain serial convertor symbol blocks into M discrete tones.

4. Subcarrier mapping: DFT output tones are mapped to the specified subcarriers for transmission. SC-FDMA systems either use localized tones or distributed spaced tones. Localized subcarrier mapping is used in LTE.

5. N-point IDFT: The transmission took place by converting mapped subcarriers back into time domain.

6. Cyclic prefix and pulse shaping: Cyclic prefix is attached to SC-FDMA symbols to deliver multipath immunity. As in the case of OFDM, pulse shaping is employed to prevent spectral regrowth.

7. RFE: It converts digital signal to analog and upconverts to RF for transmission.

In the receiver side chain, the process is essentially reversed.

3 Proposed Watermarking Model

The watermarking process can be defined as the embedding of secret data into a carrier or host signal, which later can be extracted for verification as well as validation [17, 24, 44]. It is appropriate to several applications such as temper detection, copyright protection, content authentication, and broadcast monitoring. In addition, watermarking is also used for confidentiality, reliability, and availability of the data. Image watermarking has various advantages over medical images, such as it saves bandwidth requirement, acquires small storage space, and provides confidentiality and protection against unauthenticated access to the secured data [32, 41]. The patient’s secret data are generally kept in Digital Imaging Communications in Medicine (DICOM) standard file format for any future interrogation. This private information ought to be hold secure from any alteration or tampering [18, 19, 21].

There are basically three stages involved in the watermarking of medical images. The first stage can be described as the classification of NROI and ROI, the second stage is the watermark embedding in the host image, and the last stage is the extraction of watermark information. In the classification stage, support vector machine (SVM) can play an important role, which is nothing but a machine learning algorithm that is being used widely for classification problems [5]. On the contrary, the transform domain methods such as singular value decomposition (SVD), discrete cosine transform (DCT), and discrete wavelet transform (DWT) are robust against various attacks in the embedding and extraction process [30, 31]. However, the DWT suffers from three major drawbacks (i.e. poor directional information, shift sensitivity, and lack of phase information) [8], whereas false-positive problem and higher computational cost are the main drawbacks of the SVD-based image watermarking, in which a false watermark is spotted in a watermarked image that was not originally embedded. This false-positive problem is resolved using shuffled SVD (SSVD) [42, 43]. It improves the quality of the reassembled image by partitioning an image into a set of unit images. The permuted original image by SVD with the data-independent permutation defines the SSVD [33, 34].

DWT is an advanced technique commonly used in various applications such as compression, image processing, and watermarking. In digital image watermarking techniques, DWT has its own importance due to its favorable spatial localization and multiresolution properties. Its multiresolution property reveals several information of an image [27, 29, 35]. Wavelet function partitioned the data into distinct frequency components. Discontinuities and sharp spikes of a signal can be studied in DWT, which shows its advantage over traditional transform methods [11, 28].

SVD is one of the numerical techniques used for the diagonalization of input image [45]. This numerical technique has several applications. An image I can be represented as a M×N nonnegative matrix with scalar values. SVD decomposes the M×N matrix into a diagonal matrix S of singular values and two orthogonal matrices U and V of nonnegative matrix I. The decomposition of matrix can be mathematically expressed by the given expression [7]:

I=U*S*V′

In the statistical learning theory, SVM is a new class of supervised machine learning methods, which can be used for regression or classification problems, butgenerally it is usedinclassification issues [26]. In this classification algorithm, each data item is plotted as a point in n-dimensional feature space of items with the value of a particular coordinate, and then a classification is performed to distinguish the two classes by finding the best hyperplane, as depicted in Figure 3.

Figure 3:

Support Vector Classification.

“Margin” defines the utmost breadth of two parallel lines to the hyperplane, which does not contain any data points within the parallel slabs. The data points nearby the separating hyperplane are named “support vectors”. Support vector points lie on the edge of the slab.+represents data points of category positive and – represents data points of category negative.

Medical images are more critical images, as they contain patient information for the diagnosis of the disease, which must not be distorted, while embedding the secret data into the image. This important part of the image is ROI and the other part of the image, which is not so much essential, is known as NROI. A small misclassification may cause big trouble in extracting essential information of the patient. In the statistical learning theory, SVM is a new class of machine learning method to overcome the overfitting weakness of the neural network [40]. SVM as a classifier plays an important role, which is being used widely for classification problems [5, 6, 38]. However, in medical image watermarking, SVM has not been used in transform domain for medical image classification of region so far.

In our proposed model, a double-layer security is introduced to ensure the robustness of embedded data. The embedded data are scrambled using a unique key and a transform domain-based hybrid watermarking technique is applied to embed the scrambled data into the coefficients of host image. This proposed scheme exploits the feature of spatiofrequency localization of DWT and intrinsic algebraic properties of SVD.

The soft computing technique has been applied in numerous watermarking models. However, in medical image watermarking, SVM has not been used so far. In our watermarking system, SVM is involved in the classification stage, whereas the hybrid watermarking technique (DWT-SVD) is applied in the remaining stages. The block diagram is given in Figure 4.

Figure 4:

Flow Diagram of the Proposed Watermarking Model.

The steps involved in embedding and extraction in the proposed watermarking model are discussed here.

1. Watermark embedding

   1. The NROI of host image H is decomposed into four subbands using one-level DWT.

   2. The SVD is applied to LH and HL subbands.

   3. The watermark image W is decomposed into two parts and then scrambled.

      W=P1+P2

   4. SVD is applied to the modified singular values.

      Si+αWi=UWiSWiVWi′

      To measure the robustness of the inserted watermark, a factor is used, which is known as scale factor (α).

   5. The modified DWT coefficients are obtained, i.e.

      H*i=UiSWiVi′

   6. A watermarked image is achieved by performing one-level inverse DWT.

   7. Now this watermarked image is added to the ROI pixels of image H to produce the final watermarked image W_D.

2. Watermark Extraction

   1. The watermarked image W_D is decomposed into four subbands using one-level DWT.

   2. The SVD is applied to LH₂ and HL₂ subbands of watermarked image, i.e.

      HW*i=U*iSW*iV*i′

   3. Obtain E*i=UWiSW*iVWi′

   4. Extraction of both scrambled part of watermark, i.e.

      Ew*i=(E*i−Si)/α

   5. The obtained scramble part from Step 4 is descrambled using the same secret key and then combined to form the complete extracted watermark image, i.e.

      Ew*=Ew*1+Ew*2

4 Results and Discussion

The experiment is performed using Intel Core i5 3.40 GHz processor, 1 TB HDD, 4 GB DDR3 memory, and Microsoft Windows 10. The proposed watermarking model is experimented using a JPEG grayscale image of size 256×256. The host image and watermarked image are shown in Figure 5. The database used in this experiment is taken from the MATLAB Central library. The grayscale image of size 128×128 is used as a watermark.

In the experiment, a JPEG image of size 256×256 consists of 65,536 predictor data points. These data points are further classified into two regions generally known as training data and testing data, which are used for the SVM model. For the training purpose, there are 6553 or 10% data points of the total available data, whereas the remaining 5893 or 90% data points are used to cross-validate the trained classification model. The feature values used here for classification are the intensity values and pixel positions of the input image, which are selected by the appropriate threshold value, and the result is determined by the naked eye. The proposed watermarking model and SVM classification are well discussed in Ref. [22]. The performance of our proposed watermarking model is tested with the Ganic and Eskicioglu [10] and Ramly et al. [23] methods for imperceptibility and robustness [22].

The simulation of image transmission is performed over an OFDM-based downlink physical layer. QPSK modulation is used in physical layer. BER and PSNR are calculated using the different values of SNR in the simulation at the receiver end. The average values of PSNR and BER are calculated by five trials for every SNR value as shown in Table 2. The perceptual quality is compared between the original and received images at the receiver end in Figures 5A and 6. For different SNR values, a watermark is obtained in Figure 7. Figure 7 reveals that the true recovery of the watermark data is obtained at a minimum value of SNR (i.e. 10.2 dB). The result achieved in this experiment reveals that a 10⁻⁶ BER value is realizable for a greater value of SNR (i.e. more than 10.4 dB of SNR). The PSNR of the received cover image is more than 35 dB, which is acceptable for clinical applications [13].

Several attacks are performed on the watermarked image. SSIM [39] is used for measuring the similarity between two images. The SSIM is calculated to measure the robustness of the proposed model against various attacks, such as pepper and salt noise addition, Gaussian noise, and JPEG compression, as shown in Table 3.

5 Conclusions

In the healthcare industry, medical images or information are transmitted from one place to another using wired or wireless medium. Therefore, the transmission of such information requires more security. The proposed model ensures the perceptibility and robustness of medical images against several image processing attacks. In this paper, a secured transmission of the watermarked image over a simulated LTE downlink physical layer is performed. The outcomes from this experiment reveal high-quality imperceptibility of the received image with PSNR value of 52.74 dB at SNR of 10.2 dB, which is more than the adequate value of 35 dB for visual clinical analysis. The heartening simulation outcomes can have possible applications in telemedicine.

However, the proposed model may have increased the computational complexity, which needs to be improved. In future works, it can be extended to the embedding of text and fragile watermark at different levels of transformation. Further improvement of the performance of the proposed model will be testified in the forthcoming communication.

Figure 5:

(A) Host Image and (B) Watermarked Image.

Figure 6:

Extracted Watermarked Image.

Figure 7:

Extracted Watermark at Different SNRs.