(PDF) Enhancing reliability and efficiency for real-time robust adaptive steganography using cyclic redundancy check codes

Journal of Real-Time Image Processing https://doi.org/10.1007/s11554-019-00905-7 SPECIAL ISSUE PAPER Enhancing reliability and efficiency for real‑time robust adaptive steganography using cyclic redundancy check codes Yi Zhang1 · Xiangyang Luo1 · Xiaodong Zhu1 · Zhenyu Li1 · Adrian G. Bors2 © Springer-Verlag GmbH Germany, part of Springer Nature 2019 Abstract The development of multimedia and deep learning technology bring new challenges to steganography and steganalysis techniques. Meanwhile, robust steganography, as a class of new techniques aiming to solve the problem of covert commu- nication under lossy channels, has become a new research hotspot in the field of information hiding. To improve the com- munication reliability and efficiency for current real-time robust steganography methods, a concatenated code, composed of Syndrome–Trellis codes (STC) and cyclic redundancy check (CRC) codes, is proposed in this paper. The enhanced robust adaptive steganography framework proposed is this paper is characterized by a strong error detection capability, high cod- ing efficiency, and low embedding costs. On this basis, three adaptive steganographic methods resisting JPEG compression and detection are proposed. Then, the fault tolerance of the proposed steganography methods is analyzed using the residual model of JPEG compression, thus obtaining the appropriate coding parameters. Experimental results show that the proposed methods have a significantly stronger robustness against compression, and are more difficult to be detected by statistical based steganalytic methods. Keywords  Robust steganography · STC–CRC codes · JPEG compression resistant · Statistical detection resistant 1 Introduction progress in the past few years, which brings new challenges for covert communication [17, 19]. Hence, the real-time Due to the rapid development of multimedia technology and image steganography based on instant communication tools intelligent devices, digital images processed and transmit- has become a new research hotspot in the field of informa- ted by the smart mobile devices have become an important tion-hiding techniques. However, JPEG image compression, potential carrier for covert communication [13]. Mean- which is almost always used within the mobile device image while, the deep learning technology has made a tremendous acquisition, represents a challenge to the hidden information within image content [24], such as the messages embedded by steganographic algorithms proposed in [12, 25, 27] and * Xiangyang Luo

so on. Furthermore, image steganalysis algorithms based on deep learning technology, such as Convolutional Neural Net- Yi Zhang

work (CNN) models, often use various network structures to learn the effective features of images to distinguish cover Xiaodong Zhu

and stego images [18], thus proposing higher requirements for the detection-resistant performance of steganographic Zhenyu Li

algorithms. Therefore, how to balance the resilience of embedded messages under lossy channel [1] and the detec- Adrian G. Bors

tion resistance of stego images [3] is a major problem for image steganography on mobile devices. 1 State Key Laboratory of Mathematical Engineering For detection-resistant image steganography techniques, and Advanced Computing, 62# Science Rd, Zhengzhou, current adaptive steganography algorithms, such as Highly China Undetectable steGO (HUGO) steganography [11], Wavelet- 2 Department of Computer Science, University of York, Obtained Weights (WOW) steganography [6], Minimizing YO10 5GH York, UK 13 Vol.:(0123456789) Journal of Real-Time Image Processing the Power of Optimal Detector (MiPOD) [14], JPEG UNI- STCs. For the DMAS algorithm, the embedding domain is versal Wavelet Relative Distortion (J-UNIWARD) steg- constructed based on the correspondence between quantization anography [7], and other algorithms [5], have become a tables and coefficients’ variance caused by compression, and research priority in the field of information-hiding tech- the cost function is enhanced with the side-information corre- niques. Utilizing the appropriately defined distortion func- sponding to the quantization errors of different locations, thus tions and minimizing embedding cost codes—Syndrome- realizing robust adaptive steganography based on RS-STCs Trellis Codes (STCs) [4], these algorithms can adaptively with a lower complexity. select embedding locations according to the content of Although the above three robust adaptive steganographic cover images. Consequently, such methods achieve a good methods can achieve basic resistance to JPEG compression detection resistance against steganalysis based on statistical and detection, the serious error spreading problem caused features [8]. However, these algorithms usually do not con- by STCs decoding and heavy error correction burden of RS sider the situation when the stego images are attacked during codes [23] lead to a bottleneck in the embedding efficiency the transmission through public lossy channels exposed to and detection resistance. Therefore, how to achieve higher reli- image-processing attacks, resulting in the embedded mes- ability, lower cost and higher efficiency of message coding and sages hard to survive after these attacks and the failure of embedding is a key issue in the further development of robust covert communication under lossy channels [20]. steganography. To this end, we propose to combine the cyclic In terms of JPEG compression resistance information- redundancy check (CRC) codes with STCs in this paper, to hiding techniques, robust watermarking algorithms can achieve message coding and embedding with higher reliability realize message embedding with a good visual invisibility and efficiency. The error spreading problem caused by STC and a strong resistance against JPEG compression and other decoding is also addressed in this study, thus providing a pre- image-processing attacks. Utilizing the robust embedding liminary solution for secure and reliable covert communication domains constructed based on imposing constraints [2], under lossy channel exposed to JPEG compression attacks. coefficients’ relationships [10], image features [15, 16], or In the next section, the enhanced robust steganography other methods, these algorithms can embed and retrieve based on STC–CRC codes is proposed. In Sect. 3, the fault watermarks with a high accuracy after the watermarked tolerance is analyzed, and the recommended coding param- images suffer from image-processing attacks. However, it eters are given as well. The experiment results are presented should be noted that the embedding capacity of these algo- in Sect. 4, and this paper is concluded in Sect. 5. rithms might be relatively limited considering the visual quality of watermarked images. Meanwhile, these methods often leave out the statistical detection resistance of water- 2 Enhanced robust steganography marked images, and the successful retrieval of the water- mark is not guaranteed [20], which results in a non-secure STC–CRC codes are proposed in this section to address the covert communication under the lossy channel. heavy error spreading problem caused by STCs and JPEG Since the above information-hiding algorithms cannot compression for current robust adaptive steganography meth- realize message embedding with both JPEG compression and ods. This approach considers both error detection and cor- detection resistance, utilizing the advantages of adaptive steg- rection performance, thus enhancing the framework of robust anography and robust watermarking, the DCT Coefficients adaptive steganography in terms of communication reliability Relationship-based Adaptive Steganography (DCRAS), and efficiency at the same time. Feature Region-based Adaptive Steganography (FRAS), and Dither Modulation-based Adaptive Steganography (DMAS), 2.1 STC–CRC codes is proposed in our previous studies [21, 22, 26], respectively. Utilizing the DCT coefficients’ relationship invariability The principle of the proposed concatenated code, STC–CRC against compression, the embedding domain is constructed in codes, can be described as follows. Suppose that the cover the DCRAS algorithm. Combined with the embedding cost and stego sequences with length l are c, s, respectively, and function to measure compression and detection resistance, the message sequence of length n is m. The messages are ini- messages are embedded by STCs with minimum costs after tially embedded into the cover sequences using STC codes [4], RS coding, thereby acquiring both compression and detection which can be expressed by the following formulas: resistance. On this basis, the FRAS algorithm utilizes the Har- ris–Laplacian feature to construct and select the compression 𝐬 = Emb(𝐜, 𝐦) = arg min D(𝐜, 𝐬) (1) 𝐬∈𝐂(𝐦) maintainable image regions, achieving the balance between JPEG compression and detection-resistant properties of cover 𝐦 = Ext(𝐬) = 𝐇𝐬, (2) elements. Message embedding is realized with minimal costs by combining the embedding cost function, RS coding, and 13 Journal of Real-Time Image Processing where 𝐇 ∈ {0,1}n×l is a parity-check matrix, D(c, s) is the By appropriately selecting the generator polynomial, the distortion function which can measure CRC codes can correct any random error of length 1, and { the embedding cost} of each cover element, and 𝐂(𝐦) = 𝐳 ∈ {0, 1}l |𝐇𝐳 = 𝐦 detect any burst error of length b ≤ k [9]. In the context of is the coset corresponding to syndrome m. robust embedding domain, the combination of STCs and Since the errors will spread in the extracted stego CRC codes can realize message embedding with minimum sequences after STCs decoding when stego sequences are costs while detecting and correcting the few errors caused damaged during the transmission through public lossy by compression in stego sequences. The proposed approach channels, which can be observed from Formula (2), the thus solves the error spreading problem caused by STC CRC codes [9] are utilized to encode the stego sequences decoding while acquiring higher reliability and efficiency. after STCs and detect errors in received stego sequences before STCs decoding, to guarantee the reliability of 2.2 Robust steganography framework extracted messages after JPEG compression and improve the coding efficiency of error checking and correcting Based on the structure of “Compression-resistant Domain codes. Constructing + STC–CRC Codes”, a robust adaptive steg- Suppose that the generator polynomial with the high- anography framework which can enhance both communica- est power k is G(x), and the stego sequence s with length tion reliability and efficiency is proposed in this section. The l is denoted by 𝐲 = (yl−1 , yl−2 , … , y1 , y0 ) . The CRC coding scheme of the proposed framework is shown in Fig. 1. and verification process can be presented by the following The embedding process includes the following two steps. formulas: 1. Compression-resistant domain constructing xk Y(x) + R(x) = Q(x) ⋅ G(x) (3) (a) Embedding domain and methods constructing R(x) = rk−1 xk−1 + rk−2 xk−2 + ⋯ + r1 x + r0 , (4) Construct, select, and extract the coefficients, where Y(x) = yl−1 x + yl−2 x + ⋯ + y1 x + y0 , Q(x) is the l−1 l−2 regions, or relationships that are robust to JPEG quotient polynomial, and R(x) is the remainder polynomial. compression, and apply certain embedding Then, the check sequence r and generated coding sequence method, such as that of DCRAS [21], FRAS [22] 𝐲c can be expressed as follows: and DMAS [26], to reshape cover sequence c with length lc. 𝐫 = (rk−1 , rk−2 , … , r1 , r0 ) (5) ( ) (b) Embedding cost function design Improve distor- 𝐲c = yl−1 , yl−2 , … , y1 , y0 , rk−1 , rk−2 , … , r1 , r0 . (6) tion functions of adaptive steganography consid- During the procedure of verification, if the received ering the robustness of cover elements. Design sequence is divisible by G(x), the messages are considered and calculate embedding costs corresponding to to be accurate; otherwise, it is considered that some errors the above embedding domains. have occurred during transmission. 2. Message embedding with minimized costs Fig. 1  Robust steganography framework based on “Compression-resistant Domain Construction + STC–CRC Codes” 13 Journal of Real-Time Image Processing (a) STC coding Scramble the cover sequence c and Above all, based on the proposed framework, the current robust encrypt message m to obtain 𝐜s and 𝐦e , respec- steganography methods, such as DCRAS, FRAS, and DMAS, tively. If the message length in each group of CRC can be enhanced utilizing the strong error detection capability, codes is lr , and the highest power of generator appropriate error correction capability, and high coding effi- polynomial is k, the cover length for STC coding ciency of CRC codes, thereby taking communication security, can be calculated by Eq. (1) ( ⌈∙⌉ means the ceil- reliability, and efficiency into consideration. In the following ing integer of ∙ ). Extract the first le bits from 𝐜s to sections, the enhanced version of these methods mentioned obtain cover sequence 𝐜e , and perform STC coding above are denoted as E-DCRAS, E-FRAS, and E-DMAS. to generate stego sequence 𝐬e : ⌈ ⌉ l le = lc − c ⋅ k. (7) 3 Analysis of fault tolerance lr (b) CRC coding Perform CRC coding to 𝐬e using In this section, utilizing the residual model of JPEG com- G(x) according to the group length lr , and con- pression [23], the fault tolerance of the proposed methods is nect every group of CRC codes to obtain sequence analyzed, and the coding parameters are discussed as well. 𝐫c with length k ⋅ ⌈le ∕lr ⌉ . In accordance with the Based on the similarity between JPEG compression resid- above embedding method, embed 𝐫c into the first uals and burst errors, the errors in stego images caused by k ⋅ ⌈le ∕lr ⌉ bits of the remaining cover sequence 𝐜s compression can be described as a series of Poisson points ti with length lc − le , thus obtaining stego sequence with average rate v and length d, and the number of non-zero 𝐬s . Inversely scramble 𝐬s to obtain sequence s and residuals nl in successive stegos of length l has a the Poisson generate the corresponding stego image. distribution with the following probability density function: ( ) 𝜆ns −𝜆 Accordingly, the extraction process mainly includes the P n l = ns = ns ! ⋅ e , ns = 0, 1, … (8) following two steps. where 𝜆 = v ⋅ d is the average rate of burst errors, and also 1. Compression-resistant elements’ extracting that of non-zero residuals in successive stegos of length l.   Utilizing the compression-resistant embedding Considering the error detection and correction capability domain construction methods corresponding to the of CRC codes under appropriate generator polynomials, the embedding process, such as that of DCRAS [21], fault tolerance of proposed methods is given as below: FRAS [22], and DMAS [26], extract the stego elements sequence 𝐬′ with length ls′ ( ls� = lc). 1. If 𝜆 ⋅ l ≤ k , all the errors in stego sequences of length l 2. Stego sequence decoding can be detected, thus the communication reliability can be ensured for the extracted message sequences. (a) Scrambling Scramble the stego sequence 𝐬′ to 2. If 𝜆 ⋅ l ≤ 1 , the bust error in stego sequences of length l obtain sequence 𝐬′ s , utilizing the same scrambling can be corrected; thus, the communication accuracy can algorithm as message embedding process. be guaranteed for the extracted message sequences. (b) CRC decoding Perform CRC decoding using G(x) for each group of CRC codes in the first On this basis, the recommended CRC-coding parameters le + k ⋅ ⌈le ∕lr ⌉ elements in stego sequence 𝐬′s , in are chosen corresponding to compression-resistant domains which the first le bits are the extracted STC codes with different average burst error rates 𝜆 . According to the and the rest k ⋅ ⌈le ∕lr ⌉ bits are the extracted CRC conclusion in [23], the parameter 𝜆 can be approximately codes using for error detecting and correcting. estimated by the stego sequences’ average error rates Then, obtain the STC code sequence 𝐬′e with 𝜆′ . Then, utilizing the compression-resistant embedding length le after CRC decoding. domains defined in DCRAS, FRAS, and DMAS with the (c) STC decoding Perform STC decoding to the parameters provided in Table 1, the average error rates 𝜆′ of sequence 𝐬′e utilizing the same parameters used in cover sequence constructed from randomly selected 2000 the message embedding process, and obtain the images after compression with quality factors of 65, 75, and extracted message sequence 𝐦′e . Then, decrypt 85 are calculated, and the results are shown in Table 2. message sequence 𝐦′e with the same key used to According to the error detection performance of some fre- encrypt the secret messages before embedding, quently used CRC codes [9], the appropriate coding param- and obtain extracted message sequences 𝐦′. eters can be selected for the above three compression-resistant domains corresponding to their error rates after compression, 13 Journal of Real-Time Image Processing Table 1  Parameters’ settings Parameters DCRAS FRAS DMAS Maximum cost 108 108 108 Iterations Tstep 3 ∖ ∖ Population size Ng ∖ 100 ∖ Iterations Ni ∖ 20 ∖ Quantify tables ∖ ∖ T65 , T75 , T85 (a) Table 2  Error rates of stego sequences ( ×10−4) Error rates 𝜆′ QF = 65 QF = 75 QF = 85 𝜆′ 1 of DCRAS 0.12 0.09 0.34 𝜆′ 2 of FRAS 7.26 6.73 8.02 𝜆′ 3 of DMAS 0.13 0.26 0.30 (b) Table 3  Error rates after CRC checking [9] G(x) k lr dh 𝜆 = ­10−4 𝜆 = ­10−5 0x011D 8 27 − 1 3 1.27 × 1­ 0−10 1.27 × 1­ 0−13 −11 0x080F 12 2 −1 11 4 6.97 × 1­ 0 6.97 × 1­ 0−15 −10 0x1021 16 215 − 1 4 8.78 × 1­ 0 8.78 × 1­ 0−14 (c) that is, CRC-8 (0x011D), CRC-12 (0x080F), and CRC-16 (0x1021). The error rates after error checking of different Fig. 2  One-to-one correspondence in CRC codes average error rates 𝜆 are shown in Table 3 (where G(x) is the generator polynomial, k is the highest power of G(x), lr is the code length, and dh is the Hamming distance). Table 4  Parameters used in the experiments Moreover, a one-to-one correspondence can be found Parameters Settings between the check sequence and error location when a burst error occurs, by setting the lr of the CRC codes as 127, 506, Image source BOSSbase 1.01 image database and 395, respectively, which is illustrated in Fig. 2. Image size 512 × 512 In Fig. 2, according to the location of error bit, check Quality factors 65/ 75/ 85 sequences of CRC codes in decimal are shown by different Cover sets 10, 000 × 3 colors. From the results, it can be concluded that by selecting Cover image numbers 2000 (Randomly selected)× 3 appropriate error checking and correcting codes, the errors in Payloads 0.01, 0.02, …, 0.1 bpnzAC received sequences can be found and corrected. Consequently, Secret messages Randomly generated binary sequences a good error detection performance and correction capabil- RS coding parameters (31,23) ity can be achieved. For the proposed robust steganography Compared methods DCRAS [21]/FRAS [22]/DMAS [26] framework, CRC-8 (0x011D) and CRC-12 (0c080F) are Coding parameters CRC-8 (0x011D)/ CRC-12 (0x080F) chosen for the proposed approach, because they have better Stego image numbers 2000 × 3 × 10 × (3 + 1 × 2) = 300,000 communication channel reliability and coding efficiency when compared to other coding parameters. 4.1 Experimental setting 4 Experimental results In the following experiments, the cover sets are con- In this section, the performance of the proposed methods is structed by applying JPEG compression with quality fac- tested and compared to the existing robust steganography tors of 65, 75, and 85, respectively, to the 10,000 spa- methods in terms of compression and detection resistance. tial images in the Bossbase 1.01 database. Then, 2000 13 Journal of Real-Time Image Processing images are selected randomly corresponding to each 4.2 JPEG compression resistance group of cover images with different quality factors, and the stego images are generated using DCRAS, FRAS, and After compressing the stego images generated by three exist- DMAS with (31,23) RS codes and parameters, as shown ing robust steganography methods and the proposed methods in Table 1 when considering payloads ranging from 0.01 with quality factors of 65, 75, and 85, the embedded mes- to 0.1 bpnzAC (bits per non-zero AC coefficient in DCT sages are extracted and the average error rates after CRCs’ domain). For the proposed methods, the compression- correction are calculated and are shown in Fig. 3. resistant domains are combined with STC–CRC codes, The experimental results in Fig. 3 demonstrate that when denoted as E-DCRAS/FRAS/DMAS-8/12, while the ste- comparing with current robust steganography methods, the gos are also generated under three different quality factors message extraction error rates of the proposed methods after and ten different payloads. The experiment parameters are compression are significantly reduced, especially for FRAS, shown in detail in Table 4. whose message extraction accuracy is improved by more (a) (b) (c) (d) (e) (f) (g) (h) (i) Fig. 3  Average error rates of the extracted messages 13 Journal of Real-Time Image Processing than 100 times. It is also shown in Fig. 3 that the message error rates of the proposed methods, E-DCRAS, E-FRAS, and E-DMAS, are relatively stable and at lower level when the payload varies. This outcome is because of the strong and stable error correction and detection performance of CRC codes. CRC codes are also less affected by the embed- ding domain and by the payload diversity. In addition, sim- ilar experimental results can be achieved when the stego images are compressed using JPEG with various quality factors. Thus, it can be concluded that utilizing the strong error detection and satisfactory correction capability of CRC codes, the proposed methodology achieves steganography on a communication channel robust to JPEG compression. 4.3 Statistical detection resistance (a) When using the typical Cartesian Calibrated PEV (CCPEV) and Discrete Cosine Transform Residual (DCTR) stegana- lytic algorithms, the steganalytic features are extracted from the JPEG compressed cover images considering a quality factor of 65 and the corresponding stego images generated by three robust steganography methods proposed in [21, 22] and [26], and the algorithms proposed in this paper with payloads of 0.01–0.1 bpnzAC. Then, half of the samples in the cover images and each group of stego images are selected randomly to train the ensemble classifier, while the rest are used to test the detection resistance, and the results are shown in Fig. 4. In Fig. 4, the average detection error rates of the enhanced robust steganography methods are illustrated by solid lines, and that of methods proposed in [21, 22] and [26] are illus- trated by dotted lines. From the above experimental results, (b) it can be concluded that the detection error rates against steganalytic features are increased significantly by adopting Fig. 4  Detection error rates the proposed methods. This is mainly because the high cod- ing efficiency of STC–CRC codes helps reduce the changes in cover images caused by message embedding, comparing steganography, the STC–CRC codes, with strong error detec- with the previous robust steganography methods. Therefore, tion performance, satisfactory correction capability, and low it can be concluded that, based on the improved robust steg- embedding costs, are proposed in this paper. On this basis, anography framework, the proposed methods are character- an enhanced robust steganography framework and three ized by a higher robustness against JPEG compression and a steganographic methods resisting JPEG compression and higher detection resistance against statistical features while detection through steganalysis are proposed, such that they enhancing both communication reliability and efficiency. fulfil the conditions mentioned above. The proposed meth- ods also reduce the bottleneck of the embedding efficiency and improve the detection resistance in existing robust steganography methods. To address the message extraction 5 Conclusions integrity of the proposed methods, by combining with the residual model of JPEG compression, the fault tolerance cor- In the past few years, the multimedia technology and deep responding to each embedding domain is analyzed, and the learning technology have made a tremendous progress, and recommended coding parameters are discussed. According this brings new opportunities and challenges for the covert to the experimental results, the proposed methods not only communication. For the practical demands for communica- increase robustness against JPEG compression, but also have tion reliability, efficiency, and security in real-time robust a higher steganalysis detection resistance, thus enhancing 13 Journal of Real-Time Image Processing both communication reliability and efficiency. Since the 13. Qu, Z.G., Zhu, T.C., Wang, J.W., et al.: A novel quantum stegona- proposed method only consider the JPEG compression and graphy based on brown states. Comput. Mater. Continua 56(1), 47–59 (2018) detection resistance of real-time image steganographic tech- 14. Sedighi, V., Cogranne, R., Fridrich, J.: Content-adaptive steg- niques used for covert communication, in the future work, anography by minimizing statistical detectability. IEEE Trans. we will continue to study robust steganography with resist- Inf. Forens. Secur. 11(2), 221–234 (2016) ance against multiple image-processing attacks and improve 15. Tsai, J., Huang, W., Kuo, Y.: On the selection of optimal feature region set for robust digital image watermarking. IEEE Trans. the communication security as well, thereby expanding the Image Process. 20(3), 735–743 (2011) application scenario of image steganography techniques. 16. Tsai, J., Huang, W., Kuo, Y., et al.: Joint robustness and security enhancement for feature-based image watermarking using invari- Acknowledgements  This work was supported in part by the National ant feature regions. Signal Process. 92(6), 1431–1445 (2012) Natural Science Foundation of China (NSFC nos. U1804263, 17. Wang, J.W., Ting, L., Luo, X.Y., et al.: Identifying computer gen- U1736214, U1636219, 61872448, 61772549, and 61602508), erated images based on quaternion central moments in color qua- the National Key R&D Program (nos. 2016YFB0801303, ternion wavelet domain. IEEE transactions on circuits and systems 2016QY01W0105), and the Science and Technology Innovation Tal- for video technology, pp. 1–12 (2018) ent Project of Henan Province (no. 184200510018). 18. Xu, G.S.: Deep convolutional neural network to detect J-UNI- WARD. In: ACM workshop on information hiding and multime- dia security. ACM, pp. 67–73 (2017) 19. Ye, J., Ni, J.Q., Yang, Y.: Deep learning hierarchical represen- tations for image steganalysis. IEEE Trans. Inf. Forens. Secur. References 12(11), 2545–2557 (2017) 20. Zhang, Y., Luo, X.Y., Yang, C.F., et. al.: A JPEG-compression 1. Amini, M., Ahmad, M.O., Swamy, M.N.S.: A new locally opti- resistant adaptive steganography based on relative relationship mum watermark detection using vector-based hidden markov between DCT coefficients. In: IEEE international conference on model in wavelet domain. Signal Process. 137(2017), 213–222 availability, reliability and security. IEEE, pp 461–466 (2015) (2017) 21. Zhang, Y., Luo, X.Y., Yang, C.F., et al.: A framework of adaptive 2. Bors, A.G., Pitas, I.: Image watermarking using DCT domain con- steganography resisting JPEG compression and detection. Secur. straints. In: IEEE International Conference on Image Processing. Commun. Netw. 9(15), 2957–2971 (2016) IEEE, pp. 231–234 (1996) 22. Zhang, Y., Luo, X.Y., Yang, C.F., et al.: Joint JPEG compression 3. Denemark, T., Fridrich, J.: Steganography with two JPEGs of the and detection resistant performance enhancement for adaptive same scene. In: ieee international conference on acoustics, speech steganography using feature regions selection. Multimedia Tools and signal processing. IEEE, pp. 2117–2121 (2017) Appl. 76(3), 3649–3668 (2017) 4. Filler, T., Judas, J., Fridrich, J.: Minimizing additive distortion 23. Zhang, Y., Qin, C., Zhang, W.M., et al.: On the fault-tolerant in steganography using syndrome-trellis codes. IEEE Trans. Inf. performance for a class of robust image steganography. Signal Forensics Secur. 6(3), 920–935 (2011) Process. 146, 99–111 (2018) 5. Guo, L., Ni, J., Shi, Y.Q.: An efficient JPEG steganographic 24. Zhang, Y., Ye, D.P., Gan, J.J., et al.: An image steganography scheme using uniform embedding. In: IEEE international work- algorithm based on quantization index modulation resisting scal- shop on information forensics and security. IEEE, pp. 169–174 ing attacksand statistical detection. Comput. Mater. Continua (2012) 55(1), 59–70 (2018) 6. Holub, V., Fridrich, J.: Designing steganographic distortion using 25. Zhang, Y.W., Zhang, W.M., Chen, K.J., et. al.: Adversarial exam- directional filters. In: IEEE workshop on information forensic and ples against deep neural network based steganalysis. In: ACM security. IEEE, pp. 234–239 (2012) workshop on information hiding and multimedia security. ACM, 7. Holub, V., Fridrich, J.: Digital steganography using universal pp. 67–72 (2018) distortion function. In: ACM workshop on information hiding 26. Zhang, Y., Zhu, X.D., Qin, C., et al.: Dither modulation based and multimedia security. ACM, pp. 59–68 (2013) adaptive steganography resisting JPEG compression and statistic 8. Kang, Y.H., Liu, F.L., Yang, C.F., et al.: Color image steganaly- detection. Multimedia Tools Appl. 77(14), 17913–17935 (2018) sis based on residuals of channel differences. Comput. Mater. 27. Zhou, Z.L., Yan, M., Wu, Q.M.J.: Coverless image steganogra- Continua 59(1), 315–329 (2019) phy using partial-duplicate image retrieval. Soft Comput. 2, 1–12 9. Koopman, P., Chakravarty, T.: Cyclic redundancy code (CRC) (2018) polynomial selection for embedded networks. In: International conference on dependable systems and networks. IEEE, pp. Publisher’s Note Springer Nature remains neutral with regard to 145–154 (2004) jurisdictional claims in published maps and institutional affiliations. 10. Parah, S., Sheikh, J., Loan, N., et al.: Robust and blind water- marking technique in DCT domain using inter-block coefficient differencing. Dig. Signal Process. 53, 11–24 (2016) 11. Pevný, T., Filler, T., Bas, P.: Using high-dimensional image models to perform highly undetectable steganography. In: ACM international workshop on information hiding. ACM, pp. 161– 177 (2010) 12. Qin, C., He, Z.H., Luo, X.Y., et al.: Reversible data hiding in encrypted image with separable capability and high embedding capacity. Inf. Sci. 465, 285–304 (2018) 13 Journal of Real-Time Image Processing Yi Zhang  received her B.S. and Zhenyu Li  received the B.Sc. M.S. degrees from the State Key degree in Information Comput- Laboratory of Mathematical ing Sciences from Hefei Univer- Engineering and Advanced sity of Technology, Hefei, China, Computing, Zhengzhou, China, in 2011, the M.Sc. degree in in 2014 and 2017, respectively. Computer Application Technol- She is currently pursuing the ogy from Zhengzhou Institute of Ph.D. degree at the State Key Information Science and Tech- Laboratory of Mathematical nology, Zhengzhou, China, in Engineering and Advanced 2014, and the Ph.D. degree in Computing, Zhengzhou, China. Computer Science from Univer- Her research interests are image sity of York, York, UK, in 2018. steganography and steganalysis His research interests include 3D technique. Information Hiding, Steganalysis and Machine Learning. Xiangyang Luo  received his B.S., Adrian G. Bors  received the MSc M.S., and Ph.D. degrees from degree in Electronics Engineer- the State Key Laboratory of ing from the Polytechnic Univer- Mathematical Engineering and sity of Bucharest, Romania, in Advanced Computing, Zheng- 1992, and the Ph.D. degree in zhou, China, in 2001, 2004, and Informatics from the University 2010, respectively. He is the of Thessaloniki, Greece in 1999. author or co-author of more than In 1999 he joined the Depart- 100 refereed international jour- ment of Computer Science, Uni- nal and conference papers. He is versity of York, U.K., where he currently a professor of the State is currently a lecturer. Dr. Bors Key Laboratory of Mathematical was also a Research Scientist at Engineering and Advanced Tampere University of Technol- Computing. His research inter- ogy, Finland, a Visiting Scholar ests are image steganography at the University of California at and steganalysis technique. San Diego (UCSD), and an Invited Professor at the University of Montpellier II, France. Dr. Bors Xiaodong Zhu  received his B.S. has authored and co-authored 120 research papers including 25 in jour- degree from the Department of nals. His research interests include computational intelligence, image Automation, Tsinghua Univer- processing, computer vision and pattern recognition. Dr. Bors was an sity, Beijing, China, in 2014, and associate editor of IEEE Trans. Image Processing between 2010 and M.S. degree from the State Key 2014 and of IEEE Trans. Neural Networks from 2001 to 2009. He was Laboratory of Mathematical a co-guest editor for a special issue in the Journal of Pattern Recogni- Engineering and Advanced Com- tion in 2015 and the International Journal of Computer Vision in 2018, puting, Zhengzhou, China, in respectively. Dr. Bors has been a member of the organization commit- 2017. He is currently pursuing tees for IEEE ICIP 2018, BMVC 2016, IPTA 2014, CAIP 2013 and the Ph.D. degree at the State Key IEEE ICIP 2001. He is a Senior Member of the IEEE. Laboratory of Mathematical Engineering and Advanced Com- puting, Zhengzhou, China. His research interest includes infor- mation security, embedded sys- tem security, and deep learning. 13