Robust adaptive steganography based on dither modulation and modification with re-compression
RRobust adaptive steganography based on dither modulation and modification withre-compression
Zhaoxia Yin a, ∗ , Longfei Ke a a Anhui Province Key Laboratory of Multimodal Cognitive Computation, School of Computer Science and Technology, Anhui University, 230601, P.R.China
Abstract
Traditional adaptive steganography is a technique used for covert communication with high security, but the scheme is invalid inthe case of stego image is damaged by lossy channels, such as JPEG compression of channels. To deal with such problem, robustadaptive steganography is proposed to enable the receiver to extract the secret message from the damaged stego image. Previousworks utilise reverse engineering and compression-resistant domain constructing to implement robust adaptive steganography. Inthis paper, we adopt modification with re-compression in embedding scheme to improve the robustness of stego sequences. Tobalance security and robustness, we move the embedding domain to the low frequency region to improve the security of robustadaptive steganography. In addition, we add an additional check code to further reduce the average extraction error rate basedon the framework of E-DMAS (Enhancing Dither Modulation based robust Adaptive Steganography). Compared with GMAS(Generalized dither Modulation based robust Adaptive Steganography) and E-DMAS, Experiments show that our scheme canachieve strong robustness and improve the security of robust adaptive steganography greatly when the channel quality factor isknown.
Keywords:
Robust steganography, Lossy channel, Robustness, Security
1. Introduction
Steganography is a science and technology to embeddingsecret messages in images by slightly modifying the pixel val-ues (in spatial domain) or DCT coe ffi cients (in JPEG domain)of normal images [1]. Nowadays, the most popular steganog-raphy method is adaptive steganography scheme, which definesa distortion function to calculate modification costs of all el-ements, then embedding secret messages into cover elementswith Syndrome-Trellis Codes (STCs) [2] and the correspondingcosts. Such as J-UNIWARD [3], UERD [4] in JPEG domain.The schemes [3, 4] are only suitable for laboratory environ-ment which assumes the receiver can receive the stego imagelosslessly. However, social network will execute lossy process,such as JPEG compression and resize, on the uploaded imagesbecause of limited storage and bandwidth. The stego image willbe damaged by lossy process when it is uploaded to social net-work. It is hard for the receiver to extract the secret messagefrom the damaged stego image completely. In order to applysteganography to real life, it is necessary to improve the robust-ness of adaptive steganography scheme.Currently, several studies tried to achieve robust adaptivesteganography. In GMAS [5], the robust adaptive steganogra-phy is divided into three categories according to the application:1) Upward Robust, 2) Downward Robust, 3) Matching Robust.Upward Robust is the scheme work in the quality factor ofchannel JPEG compression is not smaller than the quality fac- ∗ Corresponding author
Email address: [email protected] (Longfei Ke) tor of cover images. In this mode, Zhang et al. proposed astructure of Compression-resistant Domain Constructing + RS-STCs Codes [6]. Based on the framework, Zhang et al. pro-posed several works such as DCRAS (DCT Coe ffi cients Re-lationship based Adaptive Steganography ) [7], FRAS (Fea-ture Region based Adaptive Steganography) [8], DMAS (DitherModulation-based robust Adaptive Steganography) [9]. DCRAS[7] utilizes the robustness of the relationship between DCT co-e ffi cients to embed information. FRAS [8] obtains robust em-bedding regions based on feature region extraction and selec-tion algorithm. DMAS [9] modify middle frequency DCT co-e ffi cients based on dither modulation algorithm to embed secretmessage. Although these schemes [7–9] can achieve a high ex-traction accuracy, but weak security of the schemes is a seriousdisadvantage. Yu et al. proposed GMAS [5] based on DMAS[9]. They replace symmetric distortion with asymmetric distor-tion, combined with ternary STCs and expand the embeddingdomain to the lower frequency region. GMAS [5] can achievestrong robustness, especially when the channel quality factor Q c is known, they select images with quality factor Q c as coverimages can achieve better performance. However, GMAS [5]adopt the method of encoding secret message with RS codes,which means with the increase of payload, the security willdecline rapidly because of the large number of check codes.These schemes [5, 7–9] embed message with STC [2], but er-rors in stego sequence caused by damaged stego image will ap-pear error di ff usion phenomenon in the STCs decoding [10, 11].Based on these research, Zhang et al. proposed a new frame-work Compression-resistant Domain Constructing + STCCRCCodes [12], which adopt CRC codes [13] to encode the stego
Preprint submitted to Journal of L A TEX Templates July 17, 2020 a r X i v : . [ c s . MM ] J u l equence for less check code. Zhang et al. modify DMAS [9]with the framework and get a scheme called E-DMAS. Thescheme can achieve higher security than DMAS at high pay-loads. E-DMAS [12] can solve the problem of rapid decreasingin security caused by embedding a large number of check codes.But, the same problem as DMAS [9], secret message is embed-ded into middle frequency DCT coe ffi cients, security is still anissue, although better than DMAS [9].Downward Robust is the scheme work in the quality fac-tor of channel JPEG compression is smaller than the qualityfactor of cover images. Tao et al. proposed a robust steganogra-phy scheme for the situation in [14]. At first, they re-compressan original image with channel quality factor and get a coverimage, then they embed a secret message into the cover im-age with J-UNIWARD [3] or UERD [4] to get a stego image.The next, they modify DCT coe ffi cients of the original imageaccording to the stego image and get an intermediate image, sothat the intermediate image is re-compressed with channel qual-ity factor to obtain the stego image. The scheme can extract thesecret message from the stego image completely and achievehigh security after channel JPEG compression. But the DCTcoe ffi cients residual of the intermediate image and the originalimage are too large to guarantee security. In [15], Zhu et al. uti-lize the robustness of DCT coe ffi cient sign to implement robuststeganography.Matching Robust is the scheme which re-compress coverimages multiple times to reduce impact of social network. In[16], Zhao et al. proposed a robust steganography scheme basedon transport channel matching. They re-compress the cover im-age multiple times with the channel quality factor before em-bedding a secret message. They encode the secret messag withBCH code [17] to improve the extraction accuracy of the secretmessage. The scheme can achieve strong robustness and highsecurity. Besides, the capacity of the scheme is large. But it issuspicious behavior and time-consuming.Based on the framework proposed in [12], we mitigate thesecurity shortcomings of GMAS [5] and E-DMAS [12]. Weadopt modification with re-compression to improve the accu-racy of extracted stego sequence (robustness), which providepossibility to move the embedding domain to lower frequencyregion and pursue trade-o ff between robustness and security.The experimental results show that the proposed scheme canachieve same (or stronger) robustness and higher security (es-pecially under higher payloads) than GMAS [5] and E-DMAS[12] when the channel JPEG compression quality factor is uti-lized. CCPEV (Typical Cartesian Calibrated PEV) [18] andDCTR (Discrete Cosine Transform Residual) [19] steganaly-sis algorithms are used to extract steganalysis features of stegoimages under di ff erent payloads.The main contributions of this paper are listed as fellow:(1) Combining with the known channel compression qualityfactor, we propose a modification with re-compression basedon the framework proposed in [12] to reduce errors in the ex-tracted stego sequence.(2) We balance robustness and security by move embedding do-main to lower frequency regions to improve security.In the next section of this paper, we will introduce related works in Section 2. The proposed method will be introducedin Section 3. The experiment results are shown in Section 4.Conclusion is listed in Section 5.
2. Relate works
In this part, the basic idea of dither modulation algorithm,GMAS algorithm [5] and E-DMAS algorithm [12] are intro-duced. For more detailed information, refer to the correspond-ing papers. For convenience, we represent embedding domainswith corresponding symbols in this part.
In this paper, all matrices and vectors are represented inbold. X = ( x i j ) n × n , Y = ( y i j ) n × n represent cover image andstego image with size n × n respectively. All cover and stegoimages in this paper are JPEG images. The symbol J − ( X ) rep-resent the image that X is decompressed to spatial domain. Figure 1: Embedding domain and symbols
We represent di ff erent embedding domains in 8 × ffi cient block with di ff erent symbols, for example, E8 repre-sent the 8 DCT coe ffi cients in counter-diagonal with the samecolor as shown in Fig.1. And related embedding domain will berepresented as E number1 + number2; for example, E 78 repre-sent combination of E7 and E8. Dither modulation is an implementation of Quantization In-dex Modulation (QIM)[20, 21] watermarking scheme which isa typical digital watermark embedding scheme.Just as shown in Fig.2, according to the quantization step q , the coordinate axis is divided into multiple intervals. In thedither modulation algorithm, if a de-quantized DCT coe ffi cientvalue on coordinate axis belong to an odd interval, which repre-sents cover element 1, otherwise represent 0. Embed message1 into a de-quantized DCT coe ffi cient which belong to an eveninterval, the de-quantized DCT coe ffi cient will be modified tomiddle coordinate of the nearest odd interval, just like point p igure 2: Embedding scheme of dither modulation algorithm and modification distance is d ffi cient belong to oddinterval.In the generalized dither modulation algorithm proposed inGMAS[5], cover elements are quantized DCT coe ffi cient val-ues. When it embeds message 1 into a de-quantized DCT coef-ficient which belong to an even interval, the de-quantized DCTcoe ffi cient can be modified to middle coordinate of the twonearest odd intervals just like point p d d ffi cient belong to an odd interval. Yu et al. propose GMAS [5] with the improved embed-ding method and embedding region based on DMAS [9]. Theyachieve a wonderful trade-o ff between robustness and securitybase on the framework proposed in [6]. The basic idea of GMASwill be introduced briefly as fellow.1. For a given cover image X , calculate the de-quantized DCTcoe ffi cients of the cover image.2. Calculate embedding distortion ρ of each cover element withdistortion function (e.g. J-UNIWARD) according to Eq.1. ρ i j = (cid:88) k = n (cid:88) µ = n (cid:88) ν = | W ( k ) µν ( J − ( X )) − W ( k ) µν ( J − ( Y x ij )) || W ( k ) µν ( J − ( X )) | + σ (1)3. Calculate asymmetric distortion ρ + , ρ − according to Eq.2and Eq.3. x i j represent the de-quantized DCT coe ffi cients ofprocessed image, α ∈ [0 , ρ + i j = α · ρ i j , x i j < x ij q ij ρ i j , x i j ≥ x ij q ij (2) ρ − i j = α · ρ i j , x i j > x ij q ij ρ i j , x i j ≤ x ij q ij (3)4. Extract cover sequence C and the modification distances d + , d − from embedding domain with generalized dither modulationalgorithm. Embedding domain adopted by GMAS is E 678 asshown in Fig.1. 5. Calculate modifying costs ξ + , ξ − with modification distances d + , d − and asymmetric distortion ρ + , ρ − according to Eq.4 andEq.5. q i j is quantization step. ζ + i j = ρ + i j q i j , ξ + i j = ζ + i j × d + i j (4) ζ − i j = ρ − i j q i j , ξ − i j = ζ − i j × d − i j (5)6. Encode the massage m with RS codes to get a encoded mas-sage m (cid:48) .7. Embed the encoded message m (cid:48) into cover sequence C withternary STCs and modifying costs ξ + , ξ − to get a stego image Y . The receiver uses the same quantization table to calculatethe quantized DCT coe ffi cients, then utilize STCs decode thestego sequence after received the stego image, and then extractsecret message with RS decoding. Zhang et al. proposed a new framework of robust steganog-raphy in [12]. The basic structure is Compression-resistant Do-main Constructing + STCCRC Codes, which requires coversequence is robust. Because CRC code [13] has limited errorcorrection capability. Besides, the length of the check code isrelated to the length of the cover sequence, with the length ofthe check code increases, the probability of being detected bythe steganalysis algorithms increases, therefore, the length ofthe cover sequence should not be too long. The basic idea ofE-DMAS will be introduced briefly as fellow:1. For a given cover image X , calculate the de-quantized DCTcoe ffi cients of the cover image.2. Calculate embedding distortion ρ of each cover element withdistortion functions (e.g. J-UNIWARD) according to Eq.1.3. Extract cover sequence C and corresponding modificationdistances d from embedding domain according to dither mod-ulation algorithm. Embedding domain adopted by E-DMAS isE 78 as shown in Fig.1.4. Calculate modification costs ξ with modification distances d and distortion ρ according to Eq.6. q i j is the corresponding3 .01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 payload(bpnzac) N u m b e r errors in modified positionerrors in non-modified position Figure 3: Average number of errors in embedded modified position and non-modified position in damaged stego sequence quantization step ζ i j = ρ i j q i j , ξ i j = ζ i j × d i j (6)5. Scramble the cover sequence C and modification costs ξ toget scrambled cover sequence C (cid:48) and scrambled modificationcosts ξ (cid:48) , embed secret messages m into first l e bits of the scram-bled cover sequence C (cid:48) with STCs and obtain a stego sequence S . l e can be calculated with Eq.7, l c represent the length of thecover sequence, l r represent the length of the secret messagein each group CRC codes, k is the highest power of generatorpolynomial. l e = l c − (cid:100) l c l r (cid:101) · k (7)6. Encode stego sequence S with CRC codes [13] and thecheck code is embedded into the rest l c − l e bits scrambled coversequence and obtain a stego sequence S . Then, we can get astego sequence S which is composed of S and S .7. Inverse scramble the stego sequence S and get a sequence S (cid:48) ,modify de-quantized DCT coe ffi cients of the cover image withthe stego sequence S (cid:48) and modification distances d . Finally, wecan obtain a stego image Y .The receiver uses the same de-quantization table to calcu-late the quantized DCT coe ffi cients, then extract the stego se-quence from de-quantized DCT coe ffi cients with dither modu-lation. Extract check codes from the last l c − l e bits scrambledstego sequence with STCs decoding to correct the first l e bits.After that, extract secret message from corrected the first l e bitsscrambled stego sequence.
3. Proposed method
In this section, we analysis embedding process of GMAS[5] and E-DMAS [12], and propose ways (modification with re-compression, lower frequency embedding domain, additionalcheck code) to improve performance of robust adaptive steganog-raphy. And pseudo code of our scheme is shown in this section.
In the embedding process of GMAS [5] and E-DMAS [12],the de-quantized DCT coe ffi cients are keep unchanged whenthe corresponding quantized DCT coe ffi cients do not need tobe modified. Some of these unmodified de-quantized DCT co-e ffi cients may be in an unstable state, they may be change dur-ing the process of JPEG re-compression because of the round-ing operation and secret message embedding. To confirm thisviewpoint, we simplify the E-DMAS algorithm on embeddingdomain E 2345, we embed the secret message into the coversequence to get a stego sequence with dither modulation andSTCs, we modify cover image with the stego sequence to getstego image. Then we re-compress the stego image with chan-nel quality factor Q c to simulate channel lossy operation andextract the damaged stego sequence from the compressed stegoimage. Compare the original stego sequence and damaged stegosequence on 100 images randomly selected from BOSSbase1.01 [22]. Average number of errors of embedded modifiedposition and non-modified position in damaged stego sequenceare shown in Fig.3. Obviously, A large number of errors ap-pear in the DCT coe ffi cients which do not need to be modified,which shows that our viewpoint is correct.How to deal with these unstable DCT coe ffi cients is a keyissue. A natural approach is to adjust the unstable DCT co-e ffi cients in the stego image which called modification with4 .01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 payload(bpnzac) N u m b e r embedding without modificationembedding with modification Figure 4: Average number of errors in stego sequences of embedding without modification stage and with modification stage re-compression in this paper. The way to modify the unsta-ble DCT coe ffi cients adopted in this paper is the dither mod-ulation algorithm. Modification with re-compression processwill be described in detail in Section 3.2. In order to verify theperformance of this method, compared to simplified E-DMAS,we add a modification with re-compression process to the em-bedding stage and generate stego images with two embeddingscheme. Then, we re-compress the stego images with channelquality factor Q c to simulate channel lossy operation. Next,we extract damaged stego sequence from the compressed stegoimages and calculate average number of errors in the extractedstego sequences on 100 images randomly selected from BOSS-base 1.01 [22]. The experimental results are shown in Fig.4.The modification with re-compression process can reduce theerrors of the stego sequence and improve the robustness.Traditional adaptive steganography has proved that embed-ding message on low frequency DCT coe ffi cients is more secu-rity than high frequency. However, robustness of DCT coe ffi -cients in low frequency region is poor which has been proved inGMAS [5]. Because of the performance of modification withre-compression when the quality factor of lossy channel is uti-lized, it is possible to move the embedding domain to low fre-quency region to improve security of robust adaptive steganog-raphy. For better performance, we only consider the lower fre-quency embedding domain than GMAS [5]. Constructed em-bedding domain is obtained with experiments in Section 4.3.In the framework Compression-resistant Domain Construct-ing + STCCRC Codes [12], check code is embedded into therest l c − l e bits cover sequence without any error correction code,it will cause lots of errors in the STCs decoding stage if thereis error in the corresponding stego sequence, because of the er-ror di ff usion phenomenon of STCs. It is reasonable to encodethe rest l c − l e bits stego sequence with additional error correc-tion code. The rationality of the additional check code will be proved in Section 4.4 with experiments.It is worth mentioning that current robust adaptive steganog-raphy is not suitable for embedding with high payload due toits weak security. Therefore, a smaller embedding domain isselected to reduce the number of check codes and improve se-curity of robust steganography.Besides, considering that CRC codes have limited error cor-rection capability. We replace CRC codes with RS codes.The flowchart of our method is shown in Fig.5. First, wemodify the cover image to embed secret message according toour scheme and obtain a stego image, and then transmit thestego image to the receiver with a lossy channel. Receiver ex-tract the secret message from the stego image processed by thethe lossy channel. In the flowchart, the modification phase cor-responds to modification with re-compression. To facilitate theexplanation, we use pseudo code to introduce the steps of ourembedding scheme. In this paper, we assume that the channel quality factor Q c is known and adopt JPEG images with quality factor Q c as thecover image.1. Investigate the quality factor Q c of the lossy channel, andadopt JPEG images with quality factor Q c as cover images.2. Calculate distortion ρ of a cover image with typical distor-tion function (e.g. J-UNIWARD) according to Eq.1.3. Extract the cover sequence C and corresponding modifica-tion distances d from the constructed embedding domain withthe dither modulation algorithm.4. Calculate the modification costs ξ with the modification dis-tances d and distortion ρ according to Eq.6.5. Scramble the cover sequence C and modification costs ξ toget a sequence C (cid:48) and modification costs ξ (cid:48) in order to avoidcentralized distribution of error points. Sequence C (cid:48) is divided5 igure 5: Framework of proposed method into 3 segments( C , C , C , l C : l C : l C ≈
15 : 3 : 1 , l C represent the length of C ).6. Embedding secret messages m into the sequence C withSTCs and corresponding distortion to get a stego sequence S .7. Encode the stego sequence S with RS codes and embed thecheck code of S into sequence C with STCs and correspond-ing distortion to get a stego sequence S .8. Encode the stego sequence S with RS codes, embed thecheck code of S (called additional check code later) into se-quence C with STCs and corresponding distortion to get astego sequence S . Finally, we get a stego sequence S afterinverse scramble the stego sequence ( S , S , S ).9. Modify de-quantized DCT coe ffi cients of cover image withstego sequence S and modification distances d . We generate in-termediate image I with the modified DCT coe ffi cients.10. Re-compress the intermediate image I with quality factor Q c , and extract the stego sequence S (cid:48) from re-compressed in-termediate image I (cid:48) with the dither modulation algorithm.11. Compare S (cid:48) with S , if there is di ff erence in S (cid:48) , we mod-ify the corresponding de-quantized DCT coe ffi cients in I (cid:48) withdither modulation algorithm.12. Repeat step 10 and step 11 n times and get a final stegoimage Y .The receiver uses the same quantization table to calculatethe quantized DCT coe ffi cients, then extract the stego sequencefrom quantized DCT coe ffi cients with dither modulation algo-rithm and scramble the stego sequence S . Extract additionalcheck code from S with STCs decoding to correct the S . Sim-ilarly, extract check code from corrected S with STCs decod-ing to correct the S . After that, extract the secret message fromcorrected S sequence.
4. Experiment
In this part, we obtain the appropriate parameters of ourproposed scheme in section 3 through experiments at first. Thencompare our scheme with GMAS[5] and E-DMAS [12].
The dataset of all experiments conducted in this paper isBOSSbase 1.01[22], which contain 10000 512*512 grayscaleimages. We assume that the channel quality factors are Q c = ,
75. We compress these images with quality factor Q c . Weset the parameter h =
10 of STCs just as GMAS[5] and E-DMAS[12]. It is worth mentioning that we treat every 8-bitstego sequence as an integer and encode the integer sequencewith RS (255,251). There is a situation that errors in the stegosequence exceed the error correction ability of RS codes, it willnot be corrected. And we believe that at most one bit in eachinteger is wrong, because of the strong robustness of the coversequence. The ensemble classifier [23] is used.In the following experiment, 1000 images are selected ran-domly from BOSSbase 1.01 [22] as original images and JPEGcompress these images with quality factor Q c to get cover im-ages. We embed a secret message into a cover image with theproposed method to get a stego image, then re-compress thestego image with the quality factor of lossy channel Q c to sim-ulate channel lossy operation. We extract the secret messagefrom the re-compressed stego image and calculate a average ex-traction error rate of the secret message. We utilize the averageextraction error rate of secret message to evaluate the robust-ness of schemes.For security, we conduct experiments on whole 10000 im-ages of BOSSbase 1.01 [22]. We utilize CCPEV [18] and DCTR[19] features to extract features of cover and corresponding stegoimages. The ensemble classifier is trained with default settings.6 .2. Performance of modification with re-compression In order to verify the e ff ectiveness of modification with re-compression, we conducted a comparative experiment with Q c =
65. We perform the modification stage 0, 1, 2, 3 times andnamed recom 0, recom 1, recom 2, recom 3, correspondingly,and all the other experiment settings are the same as describedin Section 4.1, steps as shown in Section 3.2. Average extrac-tion error rates of secret message are shown in Fig.6, we cansee that with the number of modification with re-compressionincrease, the average extraction error rates of secret messagegradually decreases. Since when we perform the modificationstage more than twice, improvement in terms of robustness isnot obvious, and further perform the modification stage on theintermediate image will greatly increase the running time andrisk of being detected. Therefore, we decided to perform themodification stage twice for each intermediate image. payload(bpnzac) R e rr o r -3 recom_0recom_1recom_2recom_3 Figure 6: Average extraction error rates of after performing 0,1,2,3 times mod-ification with re-compression payload(bpnzac) R e rr o r -3 E_345E_45
Figure 7: Average extraction error rates of di ff erent embedding domain We try to improve the security of robust steganography withembedding secret message in lower frequency region. How-ever, robustness of DCT coe ffi cients in low frequency region ispoor. Fortunately, modification with re-compression can reducethe error rate of extracted stego sequence, which allows us toembed a secret message into the lower frequency region. Justas describe in section 3.1, we only consider lower frequencyembedding domain than GMAS [5]. Therefore, we conducteda comparative experiment with Q c =
65, we select E 45 as em-bedding domain at first and expand the embedding domain tothe lower frequency region gradually. All the other experimentsettings are the same as described in Section 4.1, steps as shownin Section 3.2. In Fig.7, the average extraction error rates of thesecret message are much higher when the embedding domainis E 345. Yu et al. have proved in [5] that the lower frequencythe embedding domain is, the weaker robustness the cover se-quence is. Therefore we did not further expand the embeddingdomain to lower frequency region. Embedding domain selectedin this paper is E 45. payload(bpnzac) R e rr o r -4 single check codedouble check code Figure 8: Average extraction error rates of stego sequence with sigle and doublecheck code
As described in section 3.1, since the check code of E-DMAS [12] is embedded into the rest l c − l e bits sequence withSTCs, which is not robust. To solve such problem, based onthe original framework, we propose an additional check codewhich corresponds to step 8 in section 3.2. To verify the per-formance of additional check code, we conduct an experimentwith Q c =
65, we extract the secret message from stego se-quence whether the stego sequence embedded with additionalcheck code. All the other experiment settings are the same asdescribed in Section 4.1, steps as shown in Section 3.2. Werepresent the situation without additional check code as singlecheck code, otherwise double check code. In Fig.8, the resultsdemonstrate that additional check code is useful for reducingaverage extraction error rate of secret message.7 .01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 payload(bpnzac) R e rr o r -3 GMASE-DMASproposed method (a) payload(bpnzac) R e rr o r -4 GMASE-DMASproposed method (b)Figure 9: Average extraction error rates of GMAS[5], E-DMAS[12] and proposed method with channel quality factor 65(left) and 75(right) payload(bpnzac) CC P E V E OO B GMASE-DMASproposed method (a) payload(bpnzac) D C T R E OO B GMASE-DMASproposed method (b)Figure 10: Average detection error rates of GMAS[5], E-DMAS[12] and proposed method with CCPEV(left) and DCTR(right)
In this part, we will compare our scheme with GMAS[5]and E-DMAS [12] in security and robustness. From the pre-vious experiments, we can get the final scheme. And we con-duct experiments with Q c = ,
75 correspondingly. We extractthe secret message from the stego sequence, the experiment re-sults are shown in Fig.9. Obviously, our method can achievethe same and even stronger robustness than E-DMAS [12] andGMAS[5] when payloads larger than 0.03bpnzac. Although itperforms poorly at low payloads, but not far from the compara-tive schemes.As for security, we conduct experiments on the entire Boss-base 1.01 with Q c =
65. CCPEV [18] and DCTR [19] featuresare utilized to extract features of cover and stego images. Theexperiment results are shown in the Fig.10. Compared withE-DMAS [12] and GMAS [5], our scheme greatly improve se-curity in terms of DCTR features [19]. However, improvementin resist CCPEV [18] detection ability is not obvious. The ex- periment results show that when the payloads are larger than0.03bpnzac, the resist DCTR detection ability is significantlyimproved. At low payloads, the security is not far from thecomparative schemes.
5. Conclusion
Nowadays, social networks are more and more widely usedin our lives, which provides the possibility for steganography.Because of the lossy processing of social networks, such asJPEG re-compression, the robustness of adaptive steganogra-phy needs to be improved.In this paper, we propose a scheme called modification withre-compression to improve robustness of E-DMAS[12]. Andwe move embedding domain to low frequency region to bal-ance robustness and security. In addition, we add additionalcheck codes to improve robustness. The experiment resultsdemonstrate our scheme can achieve higher security and robust-8ess than comparative schemes when payloads are larger than0.03bpnzac. The security and robustness are not far from thecomparative papers when the payload is less than 0.03bpnzac.In the future, we will look for solutions that can improvesecurity without prior knowledge of channel quality factor. In-crease embedding capacity is another future work.
Acknowledgments
This research work is partly supported by National NaturalScience Foundation of China (61872003, U1636206).
ReferencesReferences [1] J. Ye, J. Ni, Y. Yi, Deep learning hierarchical representations for imagesteganalysis, IEEE Transactions on Information Forensics and Security12 (11) (2017) 2545–2557.[2] T. Filler, J. Judas, J. Fridrich, Minimizing additive distortion in steganog-raphy using syndrome-trellis codes, IEEE Transactions on InformationForensics and Security 6 (3) (2011) 920–935.[3] V. Holub, J. Fridrich, T. Denemark, Universal distortion function forsteganography in an arbitrary domain, EURASIP Journal on InformationSecurity 2014 (1) (2014) 1–13.[4] L. Guo, J. Ni, W. Su, C. Tang, Y.-Q. Shi, Using statistical image modelfor jpeg steganography: uniform embedding revisited, IEEE Transactionson Information Forensics and Security 10 (12) (2015) 2669–2680.[5] X. Yu, K. Chen, Y. Wang, W. Li, W. Zhang, N. Yu, Robust adaptivesteganography based on generalized dither modulation and expanded em-bedding domain, Signal Processing 168 (2020) 107343.[6] Y. Zhang, X. Luo, C. Yang, D. Ye, F. Liu, A framework of adaptivesteganography resisting jpeg compression and detection, Security andCommunication Networks 9 (15) (2016) 2957–2971.[7] Y. Zhang, X. Luo, C. Yang, D. Ye, F. Liu, A jpeg-compression resistantadaptive steganography based on relative relationship between dct coe ffi -cients, in: 2015 10th International Conference on Availability, Reliabilityand Security, IEEE, 2015, pp. 461–466.[8] Y. Zhang, X. Luo, C. Yang, F. Liu, Joint jpeg compression and detectionresistant performance enhancement for adaptive steganography using fea-ture regions selection, Multimedia Tools and Applications 76 (3) (2017)3649–3668.[9] Y. Zhang, X. Zhu, C. Qin, C. Yang, X. Luo, Dither modulation basedadaptive steganography resisting jpeg compression and statistic detection,Multimedia Tools and Applications 77 (14) (2018) 17913–17935.[10] W. Liu, G. Liu, Y. Dai, Damage-resistance matrix embedding framework:the contradiction between robustness and embedding e ffi ciency, Securityand Communication Networks 8 (9) (2015) 1636–1647.[11] Y. Zhang, C. Qin, W. Zhang, F. Liu, X. Luo, On the fault-tolerant perfor-mance for a class of robust image steganography, Signal Processing 146(2018) 99–111.[12] Y. Zhang, X. Luo, X. Zhu, Z. Li, A. G. Bors, Enhancing reliability ande ffi ciency for real-time robust adaptive steganography using cyclic redun-dancy check codes, Journal of Real-Time Image Processing 17 (1) (2020)115–123.[13] P. Koopman, T. Chakravarty, Cyclic redundancy code (crc) polynomialselection for embedded networks, in: International Conference on De-pendable Systems and Networks, 2004, 2004, pp. 145–154.[14] J. Tao, S. Li, X. Zhang, Z. Wang, Towards robust image steganography,IEEE Transactions on Circuits and Systems for Video Technology 29 (2)(2019) 594–600.[15] Z. Zhu, N. Zheng, T. Qiao, M. Xu, Robust steganography by modifyingsign of dct coe ffi cients, IEEE Access 7 (2019) 168613–168628.[16] Z. Zhao, Q. Guan, H. Zhang, X. Zhao, Improving the robustness ofadaptive steganographic algorithms based on transport channel matching,IEEE Transactions on Information Forensics and Security 14 (7) (2019)1843–1856. [17] G. Forney, On decoding bch codes, IEEE Transactions on InformationTheory 11 (4) (1965) 549–557.[18] J. Kodovsk`y, J. Fridrich, Calibration revisited, in: Proceedings of the 11thACM workshop on Multimedia and security, 2009, pp. 63–74.[19] V. Holub, J. Fridrich, Low-complexity features for jpeg steganalysis us-ing undecimated dct, IEEE Transactions on Information Forensics andSecurity 10 (2) (2015) 219–228.[20] B. Chen, G. W. Wornell, Quantization index modulation: a class of prov-ably good methods for digital watermarking and information embedding,in: 2000 IEEE International Symposium on Information Theory, 2000,pp. 1423–1443.[21] H. Noda, M. Niimi, E. Kawaguchi, High-performance jpeg steganographyusing quantization index modulation in dct domain, Pattern RecognitionLetters 27 (5) (2006) 455–461.[22] P. Bas, T. Filler, T. Pevn`y, Break our steganographic system: the ins andouts of organizing boss, in: International workshop on information hid-ing, Springer, 2011, pp. 59–70.[23] J. Kodovsky, J. Fridrich, V. Holub, Ensemble classifiers for steganalysisof digital media, IEEE Transactions on Information Forensics and Secu-rity 7 (2) (2012) 432–444.cients, IEEE Access 7 (2019) 168613–168628.[16] Z. Zhao, Q. Guan, H. Zhang, X. Zhao, Improving the robustness ofadaptive steganographic algorithms based on transport channel matching,IEEE Transactions on Information Forensics and Security 14 (7) (2019)1843–1856. [17] G. Forney, On decoding bch codes, IEEE Transactions on InformationTheory 11 (4) (1965) 549–557.[18] J. Kodovsk`y, J. Fridrich, Calibration revisited, in: Proceedings of the 11thACM workshop on Multimedia and security, 2009, pp. 63–74.[19] V. Holub, J. Fridrich, Low-complexity features for jpeg steganalysis us-ing undecimated dct, IEEE Transactions on Information Forensics andSecurity 10 (2) (2015) 219–228.[20] B. Chen, G. W. Wornell, Quantization index modulation: a class of prov-ably good methods for digital watermarking and information embedding,in: 2000 IEEE International Symposium on Information Theory, 2000,pp. 1423–1443.[21] H. Noda, M. Niimi, E. Kawaguchi, High-performance jpeg steganographyusing quantization index modulation in dct domain, Pattern RecognitionLetters 27 (5) (2006) 455–461.[22] P. Bas, T. Filler, T. Pevn`y, Break our steganographic system: the ins andouts of organizing boss, in: International workshop on information hid-ing, Springer, 2011, pp. 59–70.[23] J. Kodovsky, J. Fridrich, V. Holub, Ensemble classifiers for steganalysisof digital media, IEEE Transactions on Information Forensics and Secu-rity 7 (2) (2012) 432–444.