Simple and robust watermarking scheme based on

Simple and robust watermarking scheme based on square-root-modulus technique

Chun-Yuan Hsiao, Ming-Feng Tsai & Ching-Yu Yang

Multimedia Tools and Applications An International Journal ISSN 1380-7501 Multimed Tools Appl DOI 10.1007/s11042-018-6121-3

1 23

Your article is protected by copyright and all rights are held exclusively by Springer Science+Business Media, LLC, part of Springer Nature. This e-offprint is for personal use only and shall not be self-archived in electronic repositories. If you wish to selfarchive your article, please use the accepted manuscript version for posting on your own website. You may further deposit the accepted manuscript version in any repository, provided it is only made publicly available 12 months after official publication or later and provided acknowledgement is given to the original source of publication and a link is inserted to the published article on Springer's website. The link must be accompanied by the following text: "The final publication is available at link.springer.com”.

1 23

Author's personal copy Multimed Tools Appl https://doi.org/10.1007/s11042-018-6121-3

Simple and robust watermarking scheme based on square-root-modulus technique Chun-Yuan Hsiao 1 & Ming-Feng Tsai 1 & Ching-Yu Yang 2

Received: 8 September 2017 / Revised: 16 March 2018 / Accepted: 8 May 2018 # Springer Science+Business Media, LLC, part of Springer Nature 2018

Abstract In this paper, we present a simple and robust watermarking scheme for color images. The scheme is based on the square-root-modulus technique employed in the integer wavelet domain, which allows a large number of data bits to be embedded in a host image. Simulations confirmed that marked images generated by the proposed scheme are tolerant to various attacks such as blurring, brightness, contrast, cropping, edge sharpening, inversion, JPEG/JPEG2000 compressions, noise-additions, and truncation. Additionally, the payload of the proposed method is significantly larger than that of existing watermarking techniques and the resulting perceived quality is not bad. Because the code is quite simple, it is suitable for the proposed method implemented in the mobile equipments or smart devices. Keywords Data hiding . Integer wavelet transform . Color image watermarking . Square-rootmodulus

1 Introduction Owing to the proliferation of computing techniques and the rapid development of artificial intelligence, automation and data exchange between the parties are ubiquitous around the world. Organizations are capable of performing their e-business goals on platforms comprising

* Ching-Yu Yang [email protected] Chun-Yuan Hsiao [email protected] Ming-Feng Tsai [email protected]

1

Department of Computer Science and Information Engineering, National Kaoshiung University of Applied Science, Kaohsiung, Taiwan

2

Department of Computer Science and Information Engineering, National Penghu University of Science and Technology, Magong, Taiwan

Author's personal copy Multimed Tools Appl

the internet of things, intelligent robot, cloud computing, and big data analytics. However, individuals can easily share their secret (or private) information on the Internet and data could be eavesdropped or tampered with during transmission. In addition to encryption/decryption systems, data hiding provides an economical means to solve the above issues. Generally, data hiding can be classified into two categories: steganography and digital watermarking [3, 10]. The steganographic method often provides a high data hiding capacity with imperceptibility [4, 8]. Some image steganography are irreversible, meaning that the (parts of) content of original multimedia would be damaged (or lost) at the receiver after bit extraction. To preserve the originality of valuable (or priceless) multimedia, such as law enforcement, medical and military image systems, and geographic information, researchers have designed reversible data hiding to achieve the goal [14, 23]. With the categories of the purpose, digital watermarking can be divided into robust and fragile watermarking schemes. A robust watermarking scheme often has a good performance in resisting a variety of attacks, but with limited payload [7, 13]. Their main purpose is trying to protect (or secure) hidden watermark (or message) from being damaged by adversaries as much as possible. While the primary aim of the fragile watermarking schemes focuses on the authentication of image content as well as the detection/recovery of the contents of the tampered regions [11, 12, 16]. Recently, the applications of perceptual image hashing in multimedia information security have attracted several researchers [15, 17]. Perceptual image hashing, also known as a robust image signature, can transform an input image into a compact sequence, which is capable of conducting the authentication of multimedia contents, image retrieval, and image classification. In additions, robust image signature often has good performance in anti-collision and image security. Color images are commonly used in a realistic environment, several researchers have suggested watermarking approaches to protect copyright and ownership in color images [9, 18–22]. Because our topic related to color image digital watermarking, only the related existing watermarking schemes are reviewed here. Based on radius-weighted mean and feature-embedding techniques, Yang [21] presented an efficient color image watermarking scheme. Simulations showed that the scheme is robust against various manipulations such as JPEG/JPEG200 compression, color quantization, and noise additions. The payload of the scheme is 16,200 bits with an average peak-to-signal-noise ratio (PSNR) value of 49 dB. Using the quaternion Fourier transform domain and least squares support vector machine (LS-SVM), Wang et al. [18] proposed a digital watermarking for color images. Experiments indicated that their method can resist several image processing operations and geometric attacks. However, the method only provided a payload of 4096 bits. Using the discrete cosine transform (DCT) and the difference between two DCT coefficients, Parah et al. [9] proposed a color image watermarking scheme. Simulations showed that the extracted watermarks survived various attacks while the average PSNR was high. However, the payload was 0.047 bpp (bit per pixel). Using a local quaternion polar harmonic transform (PHT) technique, Wang et al. [19] proposed an impressive color image watermarking scheme. Experiments indicated that their method is robust against common image processing operations while their average PSNR is around 47.6 dB. However, the payload of the method is only 64 bits. Since the computation time involved is long, the method cannot be used in real-time applications. Based on the fuzzy least squares support vector machine and Bessel K-form distribution, Wang et al. [20] presented a watermarking scheme for color images. Experimental results showed that the scheme is tolerant to a variety of attacks while the average PSNR is approximately 40 dB with a payload of 4096 bits. However, several limitations exist in the scheme, such as a long computation time for LS-SVM training and fragile local geometrical


distortions. Use of a quick coefficient alignment and four criteria that follow Euclidean norms of pixels, Yang [22] designed a color image watermarking scheme. Simulations confirmed that the marked images generated by the scheme were robust against various attacks. The average payload for the scheme was 130,695 bits with the PSNR being approximately 35 dB. From the above review we can see that either the payload size or computation time of existing techniques is not good enough. In this study, we propose a fast and high-capacity digital watermarking for color images. The remainder of this paper is organized as follows. Section 2 specifies the procedures of bit embedding and bit extraction, overhead analysis and overflow/underflow discussion, and the specification of the complexity and dynamicity. Section 3 presents the simulation results, and Section 4 provides the conclusions of this study.

2 Proposed method Compared with other frequency transforms, such as DCT- or DFT-based digital watermarking, the IWT-based robust watermarking schemes have the following two merits: (1) The IWTbased watermarking schemes often provide a greater hiding capacity than do watermarking schemes that are based on other frequency domains. (2) Because the procedure of forward/ inverse IWT is performed by integer operation, the computation speed is fast. In addition, existing IWT-based watermarking schemes [21, 22] have shown that the hidden bits in the IWT domain are robust against various manipulations. To obtain a high hiding storage capacity with robustness, the proposed method uses the square-root-modulus technique and embeds data bits into the low-high (LH) and high-low (HL) subbands of level 1 (L1) of the integer wavelet transform (IWT) domain. Specifically, a host image is first decomposed to the IWT domain by using the following two formulas: d j;k ¼ s j−1;2kþ1 −s j−1;2k and

s j;k ¼ s j−1;2k

d j;k ; þ 2

ð1Þ

ð2Þ

where sj, k and dj, k are the k-th low-frequency and high-frequency wavelet coefficients at the jth level, respectively [2]. The symbol ⌊x⌋ is the floor function, which is the least integer greater than or equal to x. Thereafter, two test color watermarks (each with a quarter of a host image) are embedded into the LH-/HL-subband of IWT coefficients, respectively. The details of the scheme are described in the following sections.

2.1 Bit embedding Κ−1 wrj ; wgj ; wbj j¼0 be the j-th pixel derived from the first Κ−1 input (scrambled) watermark of size Κ. Further, let C ¼ ðcrj ; cgj ; cbj j¼0 with j k j k j k pffiffiffiffiffiffi pffiffiffiffiffiffi pffiffiffiffiffiffi pffiffiffiffiffiffi pffiffiffiffiffiffi pffiffiffiffiffiffi crj ¼ wrj − wrj ; cgj ¼ wgj − wgj ; and cbj ¼ wbj − wbj be the deviation value pffiffiffiffiffiffi

pffiffiffiffiffiffi between W j and W j . These non-integer parameters have to be saved and transmitted later to an intended receiver for bit extraction. To reduce the saving space and transmission ~ ¼ ð~crj ; ~cgj ; ~cbj Κ−1 are multiplied by 10 and truncated into time, the auxiliary parameters C Without loss of generality, let W j ¼

j¼0


the integer values with ~crj ¼ crj 10 ; ~cgj ¼ cgj 10 ; and ~cbj ¼ cbj 10 : Since each ~ lies between 0 and 9, it can be represented by four bits. In other words, a deviation value in C value with a non-integer (4-byte) can be reduced from 32 bits to 4 bits. The main procedure of bit embedding is specified in the following algorithm. Algorithm 1. Hiding data bits in an RGB color image. Input: A host color image S = {(ri, gi, bi)| i = 1, 2, …, MN}, an integer (modulus value) ϕ, and a scrambled watermark W. ~ Output: A marked image S^ with auxiliary information C: Method: Step 0.

Perform L1 IWT from the host image S to obtain the coefficients H ¼ ðMN =4Þ −1 from the LH-subband of IWT coefficients. hrj; hgj; hbj j¼0

Step 1. Input a set of coefficients Hj from H. If the end of input is encountered, then proceed to Step 9. Step 2. Assign the sign marks: srj = 1 if hrj > 0, otherwise srj = − 1; sgj = 1 if hgj > 0, otherwise sgj = − 1; sbj = 1 if hbj > 0, otherwise sbj = − 1; respectively. Additionally, hrj = ∣ hrj∣ if hrjj ≤ 0; hkgj = ∣ hgjj∣ if hgjk ≤ 0; and hbj =j ∣ hbjk∣ if hbj ≤ 0. pffiffiffiffiffiffi pffiffiffiffiffiffi pffiffiffiffiffiffi wrj ; d gj ¼ wgj ; and d bj ¼ wbj : In addition, calculate Step 3. Compute d rj ¼ j k j k j k pffiffiffiffiffiffi pffiffiffiffiffiffi pffiffiffiffiffiffi ~crj ¼ wrj −d rj 10 ; ~ wgj −d gj 10 ; ~cbj ¼ wbj −d bj 10 ; and cgj ¼ save ~crj ; ~cgj ; ~cbj as auxiliary . information. . . Step 4. Evaluate hrj ¼ ðhrj −ðhrj modϕÞÞ ; hgj ¼ ðhgj −ðhgj modϕÞÞ ; and hbj ¼ ðhbj −ðhbj modϕÞÞ : ϕ ϕ ϕ Step 5. Set 1 to symbol α. Step 6. Assign the parameters: hk = hrj, dk = drj, sk = srj if α = 1; hk = hgj, dk = dgj, sk = sgj if α = 2; and hk = hbj, dk = dbj, sk = sbj if α = 3. Step 7. If dk ≥ ϕ, then do the following substeps: Step 7a. If (hk mod 2) = 1, then evaluate hk = ⌊(hk − 1) × ϕ + (dk − ϕ)⌋ × sk, else hk = ⌊hk × ϕ + (dk − ϕ)⌋ × sk. Step 7b. Assign hrj = hk if α = 1; hgj = hk if α = 2; and hbj = hk if α = 3. Step 7c. Compute α = α + 1, if α < 3 then go to Step 6, otherwise go to Step 1. Step 8. If dk < ϕ, then do the following substeps: Step 8a. If (hk mod 2) = 1, then compute hk = (hk × ϕ + dk) × sk, else hk = ⌊(hk + 1) × ϕ + dk⌋ × sk. Step 8b. Assign hrj = hk if α = 1; hgj = hk if α = 2; and hbj = hk if α = 3. Step 8c. Compute α = α + 1, if α < 3 then go to Step 6, otherwise go to Step 1. ^ Step 9. Perform inverse IWT from the marked IWT coefficients to obtain marked image S. Step 10. Stop.

2.2 Bit extraction The primary procedure of bit extraction is specified in the following algorithm. Algorithm 2. Extracting hidden bits from a marked image.


n o Input: A marked image S^ ¼ ^ri ; gî ; ^bi ji ¼ 1; 2; …; MN ; a modulus value ϕ, and ~ ¼ ð~crj ; ~cgj ; ~cbj Κ−1 : auxiliary information C j¼0 n 0 0 oΚ−1 0 0 wrj ; wgj ; wbj . Output: An extracted watermark W ¼ j¼0

Method: n oðMN =4Þ−1 ^ ^ ¼ from Step 0. Perform L1 IWT from marked image S^ to obtain H hrj; ^ hgj; ^ hbj j¼0 the LH-subband of IWT coefficients. ^ If the end of input is encountered, then proceed ^ j from H. Step 1. Input a set of coefficients H to Step 10. hgj > 0; otherwise sgj = − 1; Step 2. Set the marks: srj = 1 if ^hrj > 0; otherwise srj = − 1; sgj = 1 if ^ hbj > 0; otherwise sbj = − 1. In addition, ^ hrj ¼ j^ hrj j if ^ hrj ≤0; ^ hgj ¼ j^ hgj j if ^ hgj sbj = 1 if ^ ^bj ¼ jh^bj j if h^bj ≤0: ≤0; and h hbj modϕ: Step 3. Compute the values w0 rj ¼ ^hrj mod ϕ; w0 gj ¼ ^hgj modϕ; and w0 bj ¼ ^

Step 4. Set 1 to symbol β. ~c ~c þ1 Step 5. Compute the randomized value of ck ¼ random 10rj ; rj10 if β = 1; ck ¼ random ~cgj ~cgj þ1 ~c ~c þ1 if β = 2; and ck ¼ random 10bj ; bj10 if β = 3. [The purpose of this step is 10 ; 10 to reconstruct the values of the overhead so as to approximate their original values.] 0 0 0 Step 6. Obtain hk ¼ ^hrj ; wk ¼ wrj if β = 1; hk ¼ ^hgj ; wk ¼ wgj if β = 2; and hk ¼ ^ hbj ; wk ¼ wbj if β = 3.

Step 7. Compute a temporary value Τ ¼ hk=ϕ ; if T mod 2 = 1, then do nothing, otherwise evaluate wk = wk + ϕ. 0 0 0 Step 8. Evaluate wk = (wk + ck)2; and let wrj ¼ wk if β = 1; wgj ¼ wk if β = 2; and wbj ¼ wk if β = 3. Step 9. Calculate β = β + 1, if β < 3 then go to Step 5, otherwise go to Step 1. Step 10. Descramble and assemble the extracted bits to form the watermark W′. Step 11. Stop. Similar procedures for the proposed method to embed/extract the second input watermark into/from the HL-subband of IWT coefficients can be inferred from the above two algorithms. Notice that Steps 3 and 4 in Algorithm 1 represent the core of the square-root modulus. In Step 3, an input watermark is pixel-wised and performed by the square root operation prior to bit embedding for the next step. Because data bits are embedded in the IWT domain, the impact of the marked images can be alleviated effectively as the manipulations of third parties. In addition, the square root operation has an effect of implicit compression. Moreover, by using the modulus operations of the host pixels in Step 4, the proposed method can hide several secret bits in the host media. A flowchart of the encoding part and decoding part of the proposed method is summarized in Fig. 1.

2.3 Overhead analysis and O/U discussion As previously described, the extraction of the hidden watermark would fail at the receiver without the help of auxiliary information. Certainly, the attackers (or the third parties) are incapable of extracting hidden data bits if they have no auxiliary information at hand. The ~ ¼ ð~crj ; ~cgj ; ~cbj Κ−1 of the proposed method is associated with the size of the overhead C j¼0


Fig. 1 Flowchart of the proposed method. a Encoding part and b decoding part

input watermark. Since the overhead values in ~crj ; ~cgj ; ~cbj lie between 0 and 9, it requires 4 × 3 = 12 (extra) bits to represent the above overhead. However, to further reduce the overhead, we only use a single bit to represent each value in the set of ~crj ; ~cgj ; ~cbj : Namely, if a value of ~crj ; ~cgj ; ~cbj lies between 0 and 4, then they are quantized to 0; otherwise, they are quantized to 1. As a result, the overhead is significantly reduced from 12 to 3 bits for each set of ~crj ; ~cgj ; ~cbj . Finally, the auxiliary information can be losslessly compressed by using either the run-length coding algorithm or JBIG2 [6]. The coded result can then be sent by an out-of-band transmission to the receiver. To prevent the occurrence of overflow/underflow (O/U) from the proposed method, each restored pixel has to be checked after it is obtained by inverse IWT. Namely, prior to forming a


Input a set of coefficients Hˆ

j

Hˆ j

Hˆ j

(b) Fig. 1 (continued)

stego-pixel of the marked images (in Step 9 of Algorithm 1), the value of a restored pixel would be fixed at 255 (or 0) in case it is larger than 255 (or less than 0) (Fig. 2).

2.4 The specification of complexity and dynamicity The time complexity of the proposed method derives primarily from three parts: (i) the forward/inverse transformation of IWT of the host image; (ii) the calculations of the deviations of the input watermark; and (iii) the procedures of bit embedding/extraction in/from the LH/ HL subband. If the size of a host image and a watermark are M × N and a × b, respectively, then time complexity of the bit embedding/extraction is O(2MN) + O(2ab) + O(MN/2). Therefore, the computational complexity of the proposed method is O(N2).


(a)

(b)

Fig. 2 Two input test images. a KUAS-badge and b calligraphy painting

In addition, Algorithm 1 shows that the modulus ϕ plays a major role in our proposed method. Specifically, various PSNR and payload performances of the proposed method can be obtained by adjusting the value of ϕ. However, in our study, adjusting this value did not noticeably affect performance in terms of robustness. Because the value of each input falls within the range of 0 to 15 after the integerized square-root operation to the RGB pixels of the watermark, the value of ϕ must not be less than 8. To supplement the depiction of the characteristics of ϕ, Fig. 3 illustrates the relationship between PSNR and ϕ in the proposed method (using two color watermarks of 256 × 256 as input data). The proposed method clearly indicates that the larger the value of ϕ, the lower is the value of PSNR (or the greater is the hiding space). In addition, the PSNR performance was shown to degrade significantly as ϕ > 11. In order to achieve both the desired PSNR value and acceptable quality, the modulus ϕ with a value of 9 was employed in our proposed method.

Fig. 3 Relationship between PSNR and ϕ of the proposed method


3 Experimental results To evaluate the performance of the proposed method, several 512 × 512 color images, including two paintings, were used as host images. Each RGB pixel of the host images was represented by 24 bits, 8 bits per component. All simulations were performed on an Intel(R) Core(TM) i5-3470 CPU with 8 GB RAM. The programming language used was Python 2.7. The average execution time of the proposed data embedding algorithm was approximately 0.015 s. Two 256 × 256 RGB images, as shown in Fig. 2, were used as the test data. The integer parameter ϕ was set at 9. The marked images generated by the proposed method using modulus 9 are shown in Fig. 4. From the figures, we can see that the perceived quality is good. No apparent color distortion appeared in the resulting images. As described previously, owing to the use of square-root-modulus technique in our proposed method, two test color watermarks, namely, the KUAS-badge and the calligraphy painting, were fully embedded in the LH-/HL-subband of the IWT domain, respectively. The resultant bit rate was 2 × (256 × 256 × 24)/(512 × 512) = 12 bpp (bit per pixel). Their average PSNR was 35.70 dB, i.e., 35.53 for BLena,^ 35.74 for BBaboon,^ 36.28 for BGoldhill,^ 35.71 for BJet,^ 36.10 for BPeppers,^ 35.76 for BSailboat,^ 35.48 for BZelda,^ 35.63 for BSplash,^ 35.74 for BLegendGod,^ and 35.80 for BDoorGods.^ Moreover, Fig. 5 indicates a trade-off between the PSNR and hiding capacity of the proposed method. The payload category, represented by the x-axis, indicates that there are seven types of payloads in different sizes used: label Ba^ for the input data being two pieces of gray images of size 64 × 64, label Bb^ for two pieces of color images of size 64 × 64, and labels Bc^–Bg^ for two pieces of gray images of size 128 × 128, two pieces of color images of size 128 × 128, two pieces of gray images of size 256 × 256, a gray image of size 512 × 512, and two pieces of color images of size 256 × 256, respectively. From the figure, we can see that the PSNR values of all the test images demonstrated a similar trend as payload size increased. The PSNR was defined as follows: 2552 ; ð3Þ MSE MN ∑ ½ðri −^ri Þ 2 þ ðg i −^ gi Þ 2 þ bi −^bi 2 : Here, (ri, gi, bi) and ^ri ; gî ; ^ bi PSNR ¼ 10 log10

where MSE ¼

1 3MN

i¼1

indicate the RGB pixel values of the host image and the marked image. To demonstrate the robustness of our method, examples of extracted watermarks (after various manipulations of the marked images) are listed in this section. Two popular image processing tools—namely, Adobe Photoshop [1] and FastStone Image Viewer [5]—were also used in our demonstrations. The extracted watermark of KUAS-badge and calligraphy painting are listed in the middle and the rightmost columns of the table. The normalized correlation (NC) value is also included. The NC is defined by NC ¼ 0

0

∑ ∑ wR ði; jÞwR ði; jÞ

where

NC R ¼

i

j

∑ ∑ ½wR ði; jÞ2 i

j

ðNC R þ NC G þ NC B Þ ; 3 0

∑ ∑ wG ði; jÞwG ði; jÞ ; NC G ¼

i

j

∑ ∑ ½wG ði; jÞ2 i

j

ð4Þ ∑ ∑ wB ði; jÞwB ði; jÞ

;

and NCB ¼

i

j

∑ ∑ ½wB ði; jÞ2 i

:

Here wR(i, j),

j

wG(i, j), and wB(i, j), and wR'(i, j), wG'(i, j), and wB'(i, j), denote the RGB pixel values of the original watermark and the extracted one, respectively. From Table 1, we see that most of the extracted watermarks are recognized. Although the NC of the extracted watermark attacked by cutting off 80% from the marked image, it was still identified. We also notice that the NC of the survived watermark extracted from a marked image, which had undergone an inversion attack, is still

Author's personal copy Multimed Tools Appl Fig. 4 Marked images generated by our proposed method. a Lena, b Baboon, c Goldhill, d Jet, e Peppers, f Sailboat, g Splash, h Zelda, i LegendGod, and j DoorGods

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)


Fig. 5 Trade-off between PSNR and hiding capacity of the proposed method

recognizable. In addition, the extracted watermarks are recognized when the marked images were manipulated by compression attacks such as JPEG and JPEG2000. Similar performance is observed in the other rows of the table. From the above demonstration, we conclude that the marked images generated by the proposed method do resist attacks such as blurring, brightness, cropping, contrast, inversion, edge sharpening, JPEG, JPEG2000, noise addition, inversion, truncation, and winding. Performance comparison between our proposed method and existing watermarking schemes [9, 18–22] is specified in this section. First, Tables 2 and 3 compare our method with existing watermarking schemes [18, 20, 22] in terms of PSNR and payload (in bpp). Table 2 shows that the proposed method yields the best PSNR value among the compared methods when bpp = 0.016. It is worthy to note that the average PSNR of the proposed method is approximately 20 dB greater than that of the other two techniques [18, 20]. Table 3 shows that the average PSNR for our method was significantly higher than that derived from Yang’s technique [22] when bpp = 0.500. Notice that when the average PSNR value in our method was 35.82 dB (bpp = 12), the payload was approximately 24 times greater than that of Yang’s technique [22]. In addition, it can be seen from Fig. 5 that the average PSNR of the proposed method (40.84 dB; payload-axis labeled with Bd^), is larger than that of Wang et al.’s techniques [18, 20], while the average payload of our method is nearly 188 times larger. Moreover, the average PSNR of the proposed method (45.92 dB; payload-axis labeled with Bc^), is larger than that of Parah et al.’s scheme [9], while the hiding storage capacity of our method is nearly 3 times larger. Although the average PSNR of our method (47.19 dB; payload-axis labeled with Bb^) is slightly less than that of the other two methods, i.e., Yang’s technique [21] (48.68 dB) and Wang et al.’s approach [19] (47.60 dB), the payload size of our method is 12 times larger than that of Yang’s technique [21] and is 3750 times larger than that of Wang et al.’s approach [19]. In summary, the hiding capability provided by the proposed method is significantly larger than existing watermarking schemes.

Author's personal copy Multimed Tools Appl Table 1 Survived watermarks extracted from the marked images which underwent various manipulations

Attacks Attack free NC*=1.00 NC^= 1.00

Negative NC*= 0.74 NC^=1.00

Uniform noise (1%) NC*= 0.53 NC^= 0.90

Brightness -100% NC*= 0.61 NC^= 0.69

Edge crispening NC*= 0.77 NC^= 0.94

Cropping 80% NC*= 0.53 NC^= 0.61

Winding NC*= 0.54 NC^= 0.74

Survived Watermarked (LH) Survived Watermared (HL)

Author's personal copy Multimed Tools Appl Table 1 (continued)

Brightness +100% NC*= 0.88 NC^= 0.90

Contrast -25% NC*= 0.71 NC^= 0.78

Contrast 25% NC*= 0.13 NC^= 1.05

Truncation NC*= 0.69 NC^= 0.84

JPEG (QF=90) NC*=0.90 NC^=1.00

JPEG2000 (CR= 4.11) NC*=0.59 NC^=0.82

Gaussian Blur (0.2 pixel) NC*=0.84 NC^=0.94 a

NC for the extracted watermark from the LH-subband of IWT domain

b

NC for the extracted watermark from the HL-subband of IWT domain

c

The last 3 bits of the stego-pixels were purposely truncated

Author's personal copy Multimed Tools Appl Table 2 PSNR/hiding capacity (bpp) comparison of various methods in bpp = 0.016 Images

Lena Baboon Jet Peppers Sailboat Splash House Average a

PSNR/bpp Wang et al. [18]

Wang et al. [20]

Our method

40.24 40.31 39.54 40.07 41.01 38.98 40.34 40.07

40.42 40.20 40.56 40.55 40.42 41.78 N/Aa 40.66

57.90 60.47 60.34 60.15 60.09 59.19 57.66 59.40

Not available

To provide readers with a clearer understanding of the compared methods in terms of robustness, several types of attack are listed in Table 4. The symbol B√^ drawn in the table indicates that the compared method has the capability of resisting from the selected attack item. Namely, the extracted watermarks can survive the method as the watermarked images are manipulated by the selected attack. The table shows that the proposed method is comparable to existing methods [21, 22] in terms of robustness. Although the other four schemes [9, 18–20] perform better in terms of robustness against geometric operations attacks, they all have the drawbacks of high computation time, robustness against a limited number of attack types, and small payloads. In addition, utilizing the compared schemes of [9, 18–20] in real-time applications and in environments where a high hiding capacity is required may not be feasible.

4 Conclusion In this study, a simple and robust watermarking scheme for color images was proposed. By using the square-root-modulus technique, a large number of secret bits were effectively embedded in the host media. Experimental results confirmed that the marked images generated by our proposed method is robust against several types of attacks such as blurring, brightness, contrast, cropping, edge sharpening, inversion, JPEG/JPEG2000 compressions, noise-

Table 3 PSNR/ bpp comparison with Yang’s scheme [22] in bpp = 0.500 Images

Lena Baboon Jet Peppers Sailboat Splash House Average

PSNR/bpp Yang’s scheme [22]

Our method

36.42 27.15 34.79 34.36 32.20 38.92 41.33 35.02

43.98 45.48 44.72 45.23 45.14 43.56 43.67 44.54

Author's personal copy Multimed Tools Appl Table 4 Comparison of attack items in various methods Attacks

Blurring Brightness Contrast Cropping Equalization Gaussian (or Unif.) noise Inversion JPEG JPEG2000 RSTa (Edge) Sharpening Truncationb Winding Zigzagging

Methods Yang [21]

Wang et al. [18]

Parah et al. [9]

Wang et al. [19]

Wang et al. [20]

Yang [22]

Our method

√ √ √ √ √ √

√ N/A √ √ √ √

√ N/A N/A √ √ √

√ N/A N/A √ N/A √

√ N/A N/A N/A N/A √

√ √ √ √ √ √

√ √ √ √ √ √

√ √ √ N/A √ √ √ √

N/A √ N/A √ √ N/A N/A N/A

N/A √ N/A √ √ N/A N/A N/A

N/A √ N/A √ N/A N/A √ N/A

N/A N/A N/A √ N/A N/A √ N/A

√ √ √ N/A √ √ √ √

√ √ √ N/A √ √ √ √

a

Stands for rotation, scaling, and translation operations

b

Last three bits of the stego-pixels in the marked image were truncated

additions, truncation, and winding. Additionally, the hiding capacity of our proposed method is significantly larger than that of existing techniques. Since the computation time is short (with an average execution time of 0.015 s), it is feasible for the proposed method to be used in realtime applications.

References 1. Adobe Photoshop CC. https://www.adobe.com/products/photoshop.html 2. Calderbank AR, Daubechies I, Sweldens W, Yeo BL (1998) Wavelet transforms that map integers to integers. Appl Comput Harmon Anal 5:332–369 3. Cox IJ, Miller ML, Bloom JA, Fridrich J, Kalker T (2008) Digital watermarking and steganography, 2nd edn. Morgan Kaufmann, Burlington 4. Eielinska E, Mazurczyk W, Szczypiorski K (2014) Trends in steganography. Commun ACM 57:86–95 5. FastStone Image Viewer. http://www.faststone.org/ 6. Howard PG, Kossentini F, Martins B, Forchhammer S, Rucklidge WJ (1998) The emerging JBIG2 standard. IEEE Trans Circuits and Systems for Video Technology 8:838–848 7. Lin CC, Chang CC, Chen YH (2014) A novel SVD-based watermarking scheme for protecting rightful ownership of digital images. Journal of Information Hiding and Multimedia Signal Processing 5:124–143 8. Pan JS, Li W, Yang CS, Yan L (2015) Image steganography based on subsampling and compressive sensing. Multimedia Tools and Applications 74:9191–9105 9. Parah SA, Sheikh JA, Loan NA, Bhat GM (2016) Robust and blind watermarking technique in DCT domain using inter-block coefficient differencing. Digital Signal Processing 53:11–24 10. Phadikar A (2012) Data hiding techniques and applications specific designs. LAP LAMBERT Academic Publishing, Saarbrucken 11. Qin C, Chang CC, Chen PY (2012) Self-Embedding fragile watermarking with restoration capability based on adaptive bit allocation mechanism. Signal Process 92:1137–1150 12. Qin C, Chang CC, Chen KN (2013) Adaptive self-recovery for tampered images based on VQ indexing and inpainting. Signal Process 93:933–946 13. Qin C, Chang CC, Chen PY (2014) A novel joint data-hiding and compression scheme based on SMVQ and image painting. IEEE Trans Image Process 23:969–978

Author's personal copy Multimed Tools Appl 14. Qin C, Chang CC, Hsu TJ (2015) Reversible data hiding scheme based on exploiting modification direction with two steganographic images. Multimedia Tools and Applications 74:5861–5872 15. Qin C, Chen X, Ye D, Wang J, Sun X (2016) A novel image hashing scheme with perceptual robustness using block truncation coding. Inf Sci 361-362:84–99 16. Qin C, Ji P, Zhang X, Dong J, Wang J (2017) Fragile image watermarking with pixel-wise recovery based on overlapping embedding strategy. Signal Process 138:280–293 17. Qin C, Chen X, Luo X, Zhang X, Sun X (2018) Perceptual image hashing via dual-cross pattern encoding and salient structure detection. Inf Sci 423:284–302 18. Wang XY, Wang CP, Yang HY, Niu PP (2013) A robust blind color image watermarking in quaternion Fourier transform domain. J Syst Softw 86:255–277 19. Wang XY, Liu YN, Han MM, Yang HY (2016) Local quaternion PHT based robust color image watermarking. J Vis Commun Image Represent 38:678–394 20. Wang GP, Wang XY, Zhang CA, Xia ZQ (2017) Geometric correction based color image watermarking using fuzzy least squares support vector machine and Bessel K form distribution. Signal Process 134:197–208 21. Yang CY (2013) Robust watermarking scheme based on radius weight mean and feature-embedding technique. ETRI J 35:512–522 22. Yang CY (2017) Robust high-capacity watermarking scheme based on Euclidean norms and quick coefficient alignment. Multimedia Tools and Applications 76:1455–1477 23. Zhang X (2013) Reversible data hiding with optimal value transfer. IEEE Trans Multimedia 15:316–325

Chun-Yuan Hsiao received his B.S. and M.S. degrees in computer science and information engineering from National Taiwan University in 1999 and 2001 respectively. Then he spent half year as research assistant in the institute of information science at Academia Sinica under supervision of Dr. Chi-Jen Lu. In September of 2002 he went to Boston University, and received a Ph.D. degree in computer science in January of 2010 under supervision of Dr. Leonid Reyzin. Then he joined the computer science and information engineering department of National Kaohsiung University of Applied Sciences as an assistant professor in February of 2010. He is currently the chief of the network system section in the computer and network center in the University. Dr. Hsiao’s research interests include cryptography, watermark scheme, steganography, and network security. He is a member of the Chinese Cryptology and Information Security Association in Taiwan and served in its education promotion committee.


Ming-Feng Tsai received his B.S. degree in Computer Science and Information Engineering in 2015 from National Penghu University of Science and Technology and M.S. degree in Computer Science and Information Engineering in 2017 from National Kaohsiung University of Applied Science, Taiwan. In September 2017, he will be a software engineer at Contrel Technology. Co. Ltd., Taiwan. His recent research interests include data hiding, information security, data mining and software design.

Ching-Yu Yang received his B.S. degree in electronic engineering in 1983 from National Taiwan Institute of Technology and M.S. degree in electrical engineering in 1990 from National Cheng Kung University, Taiwan. In 1999 he received his Ph.D. degree in Computer and Information Science from National Chiao Tung University. In 1999–2005, he was a senior engineer at Chunghwa Telecom. Co. Ltd., Taiwan. He joined the Dept. of Computer Science and Information Engineering at National Penghu University of Science and Technology in February 2005, and is currently a full professor, library directory there. His recent research interests include intelligent information hiding, network security, wearable computing.