Hiding secrete data in compressed images using histogram analysis

Transcription

1 University of Woongong Research Onine University of Woongong in Dubai - Papers University of Woongong in Dubai 2 iding secrete data in compressed images using histogram anaysis Farhad Keissarian University of Woongong in Dubai, farhadk@uow.edu.au Pubication Detais Keissarian, F. 2, 'iding secrete data in compressed images using histogram anaysis', The 2nd Internationa Conference on Computer and Automation Engineering (ICCAE), 2, IEEE, Piscataway, New Jersey, USA, pp Research Onine is the open access institutiona repository for the University of Woongong. For further information contact the UOW ibrary: research-pubs@uow.edu.au

2 iding Secrete Data in Compressed Images Using istogram Anaysis Farhad Keissarian University of Woongong in Dubai Abstract In this paper, we present a data hiding scheme that embeds the secret data into the compression codes of the host image. The compression codes are associated with the visua patterns appearing in image bocks and are computed through a histogram anaysis of residua bocks of the host image. The hiding capacity for each compressed bock is not fixed and varies with its bock type. Bocks with ow visua activity hide more data than those with high visua activity. Experimenta resuts confirm that the proposed technique can provide high data capacity within the compression fie with acceptabe image quaity of the stego-images. Keywords- data hiding; image compression; bock pattern; bock histogram. I. INTRODUCTION Protection of the transmitted data over the Internet from being intercepted or tampered with has become an increasingy important issue. There are two kinds of protection techniques, cryptography and data hiding. In the former technique, the secret data are encrypted by the sender to be transferred into a set of meaningess data. Ony the ega receiver has the means to decode the meaningess data. The data hiding techniques, on the other hand, embed the secret data into a digita media caed the cover image to become a stego-image. Because the stego-image is often imperceptibe, it arouses ess attention from the maicious attackers than data encryption []. Data hiding techniques have been roughy cassified into three categories: the spatia-domain manner, the frequencydomain manner and the compression-domain manner. For the spatia-domain manner, secret data are mixed directy into the distributed pixes. The east significant bit (SB) [2] is the genera approach to hide the secret information into the SBs of each pixe of a cover image directy owing to their ower distortion in image quaity. For the frequency domain manner, the cover image must first be transformed into frequency coefficients by using a frequency-oriented mechanism such as discrete waveet transformation (DWT) [3]. ater, the secret data are combined with the reative coefficients in the frequency-form image. For the compression-domain manner, the secret data are embedded into the compression codes. In recent years, compression codes generated by we-known image compressions such as vector quantization (VQ), bock truncation coding (BTC) and simiar methods have been used for data hiding to extend the variety of cover images. The method proposed in [4], appies side match vector quantization (SVQ) with the concept of prediction to propose an adaptive data hiding scheme. In addition to the VQ compression domain, Chuang et a [5] proposed a data embedding scheme based on BTC for gray scae images. In their scheme, they predefined a threshod to cassify the type of each BTC-encoded bock as smooth or compex. Subsequenty, they embedded the secret data into the bitmap of the smooth BTC-encoded bocks. In this paper, we deveop an image hiding scheme that can hide the secrete data into compression codes of the host image, generated by the image compression technique that we reported earier in [6]. The compression codes are associated with the visua patterns appearing in image bocks and are computed through a histogram anaysis of residua bocks of the host image. The rest of the paper is organized into four sections. The concept of the proposed compression agorithm is introduced in Section 2. In section 3, the proposed hiding scheme is presented. Experimenta resuts are given in Section 4. II. IAGE COPRESSION AGORIT In the proposed image compression agorithm in [6], an image is bock coded according to the type of individua bocks. A nove cassifier, which is designed based on the histogram anaysis of residua bocks, is empoyed to cassify the bocks according to their eve of visua activity. The cassifier paces each bock into one of the two categories of uniform or edge bock. A uniform bock is coded by the bock mean, whereas an edge bock is coded by a set of parameters associated with the pattern appearing inside the bock. ike the origina BTC agorithm [7], our method encodes an edge bock by initiay computing two gray vaues and constructing a bit-map. owever, in the proposed method the computation of the gray vaues, namey the ow and high representative intensities are carried out through anaysis of the bock residuas histogram. oreover, instead of transmitting the two gray vaues, their average and difference wi be sent to decoder. Finay, instead of transmitting the whoe bit-map for the processed edge bock, an optimum bit-pattern is seected from a set of pre-defined patterns, and its index wi be transmitted. The use of these parameters at the receiver reduces the cost of reconstruction significanty and expoits the efficiency of the proposed technique. A brief description of the bock cassifier and the coding scheme are given in the next two sub-sections. A. Bock Cassifier A nove histogram-based cassification scheme has been deveoped for cassifying the image bocks [6]. The method //$26. C 2 IEEE 492

3 operates based on the distribution of the bock residuas and cassifies bock either as a ow-detai (uniform) or as a highdetai (edge) bock. The cassifier empoys the bock residuas and cassifies the bock according to their histogram. The cassification is carried out through a peak detection method on the histogram. A brief description of the cassifier is as foows: Each bock of 4x4 pixes is converted into a residua bock by subtracting the sampe mean from the origina pixes. The residua sampes are ess correated than the origina sampes within a bock. ere, two of the most important oca characteristics of the image bock are considered: centra tendency, represented by the mean vaue and the dispersion of the bock sampes about the mean, which is represented by the residua vaues. The chaenge here is to anayze the dispersion of the residua vaues about the mean. One way of achieving this is to sort the histogram of the bock residua sampes. As the neighboring pixes in the origina bock are highy correated, the residua sampes wi tend to concentrate around zero. One can then quantize the residua sampes prior to forming the histogram. The histogram of the quantized residuas may then be formed and anayzed by simpy detecting its peaks. Based on the distribution of the residua sampes within the test images, we choose to appy a coarse quantization, in particuar a 5-eve non-uniform quantizer. We now define q j as the output of the quantizer with index j, as shown in Fig.. The histogram of the quantized vaues h ( q j ) may then be formed to provide the occurrence of q j. The quantized residua histogram (QR) is then anayzed by simpy detecting its peaks. According to the number of detected distinct peaks on the histogram, image bocks can be paced into two major categories of uniform and edge bocks. A histogram with a unique peak at its centre (unimoda histogram) identifies a uniform bock. Whereas, the existence of two distinct peaks impies that the processed bock is an edge bock. Fig. 2 and Fig. 3 show exampes of both types of bocks. A peak on the histogram indicates a high score of residua vaues; therefore it is fair to concude that there is a considerabe number of pixes that have the same dispersion about the bock mean. This, in turn wi ead us to concude that the gray eve vaues of these pixes are very cose to one another. ence, this group of pixes can be represented by a singe gray vaue. In this anaysis, a distinct peak on the QR of the processed bock represents a gray vaue j, given as : j = mean + q j () (e) (e) Figure 2. (a) Origina uniform bock; sampe mean = 8, (b) Residuas (c) quantized residuas, (d) quantizer s indexes, (e) bock QR Where mean is the bock mean. For a uniform bock, since the singe peak occurs at the center of the histogram, where q j =, then from Eq. the representative intensity j wi be the same as the bock mean. For an edge bock, the two peaks of the QR, which are positioned on the eft and right hand side of the centre (j=) represent the ow representative intensity and the high representative intensity, respectivey. If the two peaks are positioned at indexes j and j, the two representative intensities are cacuated as: = mean + q j In Fig. 3, using Eq. 2 ; 53. mean = 25, and = mean + q j (2) and = 25 + (-39) = 86 and (a) are computed = = (b) Figure. The quantizer output with index j (c) (d) 493

4 (e) Figure 3. (a) Origina edge bock: sampe mean = 25 (b) Residuas (c) quantized residuas, (d) quantizer s indexes, (e) QR B. Coding of Image Bocks Once the image bocks have been cassified, the coder, switches between a one-eve (uniform bock) and a bieve(edge bock) representation. A uniform bock is encoded by transmitting the bock mean pus an indicator to inform the decoder that the bock is uniform. By forciby custering a pixes in an edge bock into two groups, a bi-eve approximation of the bock is obtained. The custering partitions a bock W into two sets of pixes, W andw, such that W = W W and W W = Φ. The custering is carried out by marking the pixes of set W and W by and, respectivey. Thus the custering can be represented as a bit-pattern, B = { b, b 2,... b 6 bi (,)}. By seecting the bock mean as a threshod, the bit-pattern can be generated as : if x > i mean b = (3) i if x i mean where, xi W are the intensities of the pixes of the edge bock. It is noted that, and are the representative intensities of the set W and W, respectivey. Figure. 4: Set of 32 pre-defined patterns. ike the origina BTC, an edge bock can be coded by transmitting the representative intensities and the bit-pattern. owever, in our method, we transmit the average, and difference, of the representative intensities, defined by : + = 2 = (4) 2 The vaues and represent the ow and high frequency components, respectivey. It is evident fro eq.3 that = + and =. During the reconstruction, the coded bock can be constructed by : + if b W i b = (5) i if b W i It shoud be noted that for a uniform bock, since both representative intensities are the same as the bock mean, therefore, = mean and =. Instead of transmitting the whoe bit-pattern of an edge bock, further bit reduction can be achieved by finding the best match for the bock bit pattern from a set of pre-defined patterns, P k, k =,,2,, N. A set of 32 patterns shown in Fig. 4, which preserve the ocation and poarity of edges in four major directions and their compements making N=64 is used in our method. The pattern matching stage is carried out by performing a ogica excusive NOR operation on the bock bit-pattern and each pattern from the set to cacuate a matching score, ms, given as : = 3 = 3 ms ( P b ) i j ij ij (6) The pattern with the highest ms is seected and its index k wi be transmitted. Since, the proposed method sends k 64 instead of the whoe bock bit-pattern, ony og 2 = 6 bits are transmitted. Each image bock is therefore encoded by generating a tripe (,, k). It shoud be aso noted that, since for a uniform bock, no pattern index is transmitted, therefore the compressed code for such a bock is the pair (, ), where =. The vaue in the tripe (,, k), can be coded by 8 bits, whereas coding requires ony 6 bits, as its standard deviation is smaer than that of and. Therefore, the compression code of an edge bock requires 2=(8+6+6) bits to be transmitted. For a uniform bock, the number of bits required to code the pair (,) is 9=8+. III. PROPOSED DATA IDING SCEE The proposed hiding scheme embeds the secret data into the compression codes of the host image, generated by the proposed compression agorithm, described in the previous 494

5 section. The secret data is a random bit stream of and. For the sake of data security, secrete data shoud be encrypted by a secrete key before being embedded. After embedding secrete data into the compression codes, one can use the decompression procedure to obtain the stego-image. The data embedding and data extraction phases are described in the foowing sub-sections. A. Data Embedding Phase In this phase, the encoder receives the secret bits to hide them in the compression codes of uniform and edge bocks. The hiding capacity for each compressed bock is not fixed and varies with its bock type. For a uniform bock, and the bit-pattern of the bock are used for data embedding. Whereas, an edge bocks uses and for hiding the secrete bits. Since, the pixe intensities in uniform bocks are cose to their neighboring pixes, even though the bits in their bitpatterns are changed, the reconstructed pixe vaue is sti cose to its origina one. ence, a the bits in the bit-pattern can be repaced by secret bits. In the proposed compression agorithm, no bit-pattern or pattern index is transmitted whie coding a uniform bock. This significanty contributes to a high compression ratio. owever, in the proposed data hiding method, the compression code of a uniform bock is enhanced to incude the bock bit-pattern. The enhanced compression code of a uniform bock is then the tripe (,, B), where B is the bitpattern of the bock. This resuts in significanty enarging the hiding capacity at the cost of ower compression ratio. Athough, the bit-pattern is transmitted for a uniform bock, but the reconstruction is sti carried out by repacing a the eements of the bit-pattern by the bock mean. Therefore, the quaity of the reconstructed uniform bock wi be the same, if the bit-pattern is not transmitted. In the proposed method, a uniform bock has the capacity of embedding sixteen bits into its bit-pattern and two bits in its vaue. To further enarge the hiding capacity of the host image whie retaining good image quaity, the proposed scheme aso hides two extra bits in each of and vaues of an edge bock. The modification of and for embedding the secrete bits is as foows. The encoder receives two secrete bits and then transform these two bits into a secret decima digit sd. Next, the encoder uses the and vaues to compute F given as : F = mod 5 F = mod 5 (8) et sd and sd denote the secret digits to be hidden in and, respectivey. The modified vaues and are then computed as : = + ( sd F ) = + ( sd F ) (9) From Eq. 9, it is evident that no modification is needed if a secrete digit is equa to the cacuated vaue F. The embedding phase of the proposed method foows the steps isted beow : Step : Obtain the tripe compression code for the processed bock. If the bock is uniform, go to step3, otherwise go to step 2. Step 2: Embed 2 bits in to get and 2 bits in to get. Transmit the tripe (,, k). Go to step 4 Step 3: Embed 2 bits in to get, and 6 bits in B to get B. Transmit the tripe (,, B ). Go to step 4. Step 4: Repeat step, unti the entire bit stream is embedded. Step5 : Decrypt the extracted information to obtain the origina secrete data B. Extraction Phase The data extraction phase is simiar to the data embedding phase. The decoder receives the tripe code for each bock. If the bock is uniform, 2 bits are retrieved from and 6 bits from the transmitted bit-pattern B. On the other hand, if the received code is for an edge bock, 2 bits are extracted from and another 2 bits are extracted from. The secret digit sd can be extracted by computing sd = mod 5 and sd = mod 5. The data extracting process is repeatedy executed unti a the secret bits are retrieved. The extracting phase of the proposed method foows the steps isted beow : Step : Obtain the tripe compressed code for the bock. Go to step 2. Step 2: If =, the tripe is (,, B ). Extract 2 bits from and 6 bits from B, and go to step 4, otherwise go to step 3. Step 3: The tripe is (,, k) Extract 2 bits from and 2 bits from. Go to step 4 Step 4 : Repeat step, unti a the secret bits are retrieved. IV. SIUATION RESUTS We have evauated the performance of the proposed coding scheme through a computer simuation on two grayeve images ena and Pepper, shown in Fig.5a and Fig. 6a. These images are 8 bits per pixe and 52 x 52 pixes in size. The simuation patform is icrosoft Windows P, Pentium III, and the proposed scheme is impemented using atab. Three performance matrices are used to measure the performance of the proposed compression and hiding schemes : iding capacity (CE), compression ratio (bpp), and image quaity (PSNR). Tabes and 2 show the compression resuts, the hiding capacity, C and the image quaity (PSNR) for the cover images ena and Pepper. TABE I Image P(u) CR (bpp) TC (Kbyte) C (Kbyte) PSNR (db) ena 63% Pepper 72% P(u): Popuation of uniform bocks 495

6 CR : Compression Ratio, TC: Transmitted Code C : iding Capacity From Tabe I, it can be observed that more than haf of the transmitted code is used to embed data, yet keeping a satisfactory image quaity. To further investigate the performance of the proposed scheme, we compare the stegoimage quaity and the hiding capacity of the proposed scheme with Chuang et a. s scheme for a threshod of T=5. The resuts are shown in Tabe II. From Tabe II, we observe that the proposed method outperforms Chuang et a. s scheme in the hiding capacity by 5%, with a very itte distortion in the quaity. TABE II Image ethod PSNR (db) iding Capacity (Kbyte) ena Chuang et a Proposed Pepper Chuang et a Proposed Compression Ratio (bpp) From Tabe II, we observe that the proposed method outperforms Chuang et a. s scheme in the hiding capacity by 5%, with a very itte distortion in the quaity. Since the REFERENCES [] uuu Petitcoas, F.A.P., Anderson, R.J., and Kuhn,.G.: Information hiding a survey, Proc. IEEE, 999, vo. 87, (7), pp [2] J. CK. Chan and. Cheng, iding data in images by simpe SB Substitution, Pattern Recog., vo. 37(3), pp , 24 [3] P. Bao and. a, Image adaptive watermarking using waveet domain singuar vaue decomposition, IEEE Trans. Circuits Syst. Video Techno., 25, vo. 5, (), pp [4] SC. Shie, SD. in and C. Fang, Adaptive data hiding based on SVQ prediction, IEICE Trans Inform Syst., vo. E89-D(), pp [5] JC. Chuang and CC. Chang, Using a simpe and fast image compression agorithm to hide secret information. Internationa Journa of Computes and Appications, vo.28 (4), pp , 26 compression agorithm described in section II, does not transmit the bit-pattern of a uniform bock B, we aso considered hiding data into the compression code of a uniform bock without transmitting the bock bit-pattern, B. As mentioned earier, since a uniform bock is reconstructed by repacing a the bock pixes by the bock mean, the quaity of reconstructed bock wi be the same in either cases. Tabe 3 shows the resuts for this case. TABE III Image P(u) CR TC C PSNR (bpp) (Kbyte) (Kbyte) (db) ena 63% Pepper 72% By comparing the resuts in Tabe I and Tabe II, we note that non-transmission of bit-patterns of uniform bocks resuts in doubing the compression ratio, at the cost of dramatic reduction of hiding capacity by 8%. Depending on the amount of secrete data, one can seect the suitabe method. Fig. 5 and Fig. 6 show the origina test images and the compressed stego-images for both cases. [6] F Keissarian, Nove quad-tree predictive image coding technique using pattern-based cassification, Proc. SPIE, Visua Communications and Image Processing (VCIP-23), vo. 55, pp , June 23, ugano, Switzerand. [7] E.J. Dep and O.R. itche, Image compression using bock truncation coding, IEEE Trans. Commun., vo. 27, pp , 979. (a) (b) (c) (a) (b) (c) Figure 5. The compression and embedding resuts for the Image of ena : a) Origina image, b) Compressed stego-image: C=26.76 Kb, at.46 bpp, c) Compressed stego-image: C=5.5 Kb, at.8 bpp. Figure 6. The compression and embedding resuts for the Image of Pepper : a) Origina image, b) Compressed stego-image: C=28.6 Kb, at.48 bpp, c) Compressed stego-image: C=5. Kb, at.75 bpp. 496