Keywords

1 Introduction

Enormous usage of cloud computing-based systems and transmission of massive multimedia information over the Internet have created a prominent need for state-of-the-art information security frameworks [1]. Steganography, cryptography, and digital watermarking are the methods developed over the years to ensure secure and seamless transmission of data from transmitter to receiver. With the aid of digital watermarking, it is feasible to protect the ownership of any multimedia content and to defend intruders against any sort of attack [2,3,4]. Digital watermarking can be mostly classified into three domains, namely frequency domain, spatial domain, and hybrid domain. Hybrid and frequency domain watermarkings are robust and have high computational complexity, whereas watermarking in the spatial domain has added advantages of simplicity and low computational complexity [2]. However, nowadays techniques in wavelet domain are becoming more popular because of the accurate human visual system (HVS) modelling [5]. Recently, outstanding progress has been made in the area of digital image watermarking. Nayak et al. proposed a visual saliency-based LSB replacement method, which has the advantage of increased imperceptibility but at the cost of less robustness [3]. Basu et al. proposed another visual attention model-based pixel adaptive LSB replacement method, which is good in terms of robustness but has a minimal peak signal-to-noise ratio (PSNR) [2]. Verma et al. proposed a novel watermarking scheme based on dynamic stochastic resonance (DSR) and lifting wavelet transform (LWT). In this technique, the attacked image is subjected to DSR to enhance the watermark coefficients to achieve maximum correlation [6]. Tian et al. gave another HVS-based approach of the unique dual watermarking for simultaneous copyright protection and tamper detection [4]. This technique is very robust for all kinds of single and combined attacks but has higher computational complexity. We have incorporated the integer-to-integer LWT instead of conventional discrete wavelet transform (DWT) to segregate the low-frequency component of the image, and then the binary watermark is infused inside the cover image via dynamic LSB replacement (DLSBR). However, only watermarking is not enough to protect the ownership information inside an image. The watermark logo must be encrypted with some sort of encryption in order to attain added security [7]. Therefore, the original watermark logo is transformed into a random chaotic image using symmetric key cryptography (SKC) based on a private key. The same private key is needed at the time of decryption as used for encryption. The rest of the paper contains some other sections like proposed methodology followed by result analysis and conclusion.

2 Proposed Methodology

2.1 Lifting Wavelet Transform

The LWT proposed by W. Sweldens is the second-generation wavelet transform, an advancement of DWT [8]. It is faster and requires lesser memory space as compared to the traditional wavelet [9]. It generates integer coefficients for all sub-bands unlike the conventional first-generation scheme, which generates floating point coefficients, which may be truncated at any point of time during further processing [9]. LWT transforms the original image to four sub-bands, namely LL, containing low-frequency components, and most of the image energy and \(LH, \ HL, \ and \ HH\) comprising of high-frequency detailed information. Therefore, we use LL sub-band to embed the watermark data. Lifting scheme in general consists of three steps, namely divide, dual lifting, and primal lifting (Fig. 1).

  1. 1.

    Divide Operation: The sample image \(X(i,\ j)\) is divided into two subsets, i.e. even sample set \(X_e\left( i,\ j\right) \) and odd sample set \(X_o\left( i,\ j\right) \) as described in Eqs. 1, and 2.

    $$\begin{aligned} X_e(i, j) = X(i, 2j) \end{aligned}$$
    (1)
    $$\begin{aligned} X_o(i, j) = X(i, 2j+1) \end{aligned}$$
    (2)
  2. 2.

    Dual Lifting: The odd sample set \(X_o(i,\ j)\) is estimated from the local neighbourhood even coefficients with a prediction operator P. The error in odd sample prediction is used to generate the high-frequency coefficients h(i, j) as described in Eq. 3.

    $$\begin{aligned} h(i, j) = X_o(i, j) - P[X_e(i, j)] \end{aligned}$$
    (3)

From Eq. 4, we can recover the odd sample set as shown below

$$\begin{aligned} X_o(i, j) = h(i, j) + P[X_e(i, j)] \end{aligned}$$
(4)
  1. 3.

    Primal Lifting: The low-frequency coefficients \(l(i,\ j)\) are generated by revising the even sample with the updating value \(U_h(i,j)\) as shown in Eq. 5.

$$\begin{aligned} l(i,j) = X_e(i, j) + U_h(i, j) \end{aligned}$$
(5)
Fig. 1
figure 1

Lifting wavelet transform

Full size image

2.2 Symmetric Key Cryptography

The robustness of the proposed scheme is further enhanced by addition encryption to the watermark through symmetric key cryptography, which covers the inner information from eavesdroppers. The encryption process includes the following steps

  1. 1.

    Segregate all rows of the binary watermark in separate arrays.

  2. 2.

    Convert the user-given key to 16-bit binary sequence.

  3. 3.

    Perform bit-wise XOR between the 16-bit binary sequence and the segregated rows of the watermark.

  4. 4.

    Perform Step 3 until all the rows are processed.

  5. 5.

    Concatenate all the processed rows to get the encrypted watermark.

The decryption process is exactly the same as the encryption process, as performing XOR again between the segregated rows and the binary version of the user-given decryption key will yield the original unprocessed rows. Figure 2 depicts the original as well as the encrypted watermark.

Fig. 2
figure 2

Watermark before and after encryption

Full size image

2.3 Watermark Encoder and Decoder

Let I be the original host image of size \(m \times n\). The image is represented as

$$\begin{aligned} I=X(i,j) \ | \ 0 \le I<m,\ 0\le I<n,\ X(i,j)\in \{0,\ 1,\ 2,\ {\mathbf \dots },\ 255\} \end{aligned}$$
(6)

Also, W is the original binary logo of size \(p \times q\) and characterized as

$$\begin{aligned} W=V(i,j)\ |\ 0\le I<p,\ 0\le I<q,\ V(i,j)\in \{0,\ 1\} \end{aligned}$$
(7)

After encryption, the original logo W is changed to \(W_k\) and represented as

$$\begin{aligned} W_k=A_{\text {enc}}(W,k) \end{aligned}$$
(8)

where \(A_{\text {enc}}\) is the encryption operator and k is the private encryption key. The decrypted watermark \(W_d\) is represented as

$$\begin{aligned} W_d=A_{\text {dec}}(W_d,k) \end{aligned}$$
(9)

where \(A_{\text {dec}}\) is the decryption operator.

Fig. 3
figure 3

Watermark encoder and decoder block diagram

Full size image

Figure 3 represents the watermark encoding and decoding procedures. Here, the ‘approximation’ or the ‘LL’ component is isolated from the image via LWT. After the watermark encryption process, based on the LL coefficient value each watermark bit is infused inside the LL block of the cover image via a novel method called the dynamic least significant bit replacement (DLSBR). In DLSBR, the LSBs of the LL block coefficients are replaced dynamically with the watermark bits depending upon the LL coefficient values. If the LL coefficient value is less than 100 then last 3 bits, if coefficient value is less than 200 and greater than 100 then last 2 bits, otherwise last 1 bit of the LL coefficient is replaced with the watermark bit. This dynamic substitution has proven to be more robust than conventional single-bit replacement. Being a blind watermarking process, the decoder does not need the original host image in the time of decoding. The decoding process is depicted in Fig. 3b. In the decoding process, each watermark bit is extracted from the LL coefficients using dynamic LSB extraction. To decrypt the extracted watermark, one should provide the same private key as used in the time of encryption. However, the robustness against attacks and the visual imperceptibility of the proposed algorithm must be evaluated using some well-established quality metrics, which is the topic of discussion in the next subsection.

3 Result Analysis and Discussion

To test the imperceptibility and the robustness of the scheme proposed, we have taken several standard \(512\times 512\) grayscale and colour images and one binary watermark of dimension \(10\times 10\). Figure 4 depicts four out of the various cover images we have used. Figure 5 shows the corresponding watermarked image along with corresponding PSNR and structural similarity index (SSIM).

Fig. 4
figure 4

Original cover images

Full size image
Fig. 5
figure 5

Watermarked images and their PSNR and SSIM values

Full size image
Table 1 Result of various imperceptibility metrics on Lena image
Full size table

From Fig. 5, it is evident that the watermark image has no visible hint of the hidden logo and also all of the images possess high PSNR and SSIM value. To prove the imperceptibility of the proposed scheme, Table 1 has been presented with some renowned quality metrics [2]. Metric values in S. No. 1–2, 5–9, 12–15 depict premier quality as higher metric value denotes better imperceptibility [2]. On the other hand, metric values in S. No. 3–4, 10–11, also illustrate that the hidden logo is completely invisible as lower metric value denotes higher imperceptibility [2]. Hence, the proposed algorithm is proven to be beneficial in every test of imperceptibility. ‘Robustness’ deals with the proficiency of the watermarking algorithm in correctly detecting the watermark in the cover image after common geometric and signal processing attacks [1]. Figure 6 is presented with some attacked watermarked images along with the corresponding retrieved watermarks. It is evident from Fig. 6 that the watermarked image can sustain the watermark even after strong signal processing and geometric impairments.

Fig. 6
figure 6

Attacked watermarked images and corresponding extracted watermarks

Full size image
Table 2 Imperceptibility comparison with various state-of-the-art methods
Full size table

Normalized cross-correlation (NCC) has been taken as the primary metric to test our approach. In most of the cases, we got the NCC value of 1 or almost 1. Other metric values like SSIM, MSE, PSNR are also in support of the superior ruggedness of the algorithm. To establish the superiority of the proposed framework, a competitive analogy with various state-of-the-art schemes has been drawn in Table 2. It can be observed that, among all models, the projected approach gives highest PSNR as well as SSIM.

4 Conclusion

A wavelet domain, blind watermarking, and secure watermarking framework have been presented in this paper. This scheme impregnates the watermark by substituting the LSBs of the LL block coefficients with the watermark bits. Furthermore, the original watermark logo is encrypted with SKC for added security. Experimental results establish the fact that the projected approach has excellent imperceptibility and is robust against most forms of signal processing impairments. In near future, we aim to make the algorithm faster for effective hardware implementation.