Secure Steganography in Compressed Video Bitstreams

Secure Steganography in Compressed Video Bitstreams

Bin Liu, Fenlin Liu, Chunfang Yang and Yifeng Sun Zhengzhou Information Science and Technology Institute,

Zhengzhou, P. R. China 450002 [email protected], [email protected]

Abstract

A new compressed video secure steganography (CVSS) algorithm is proposed. In the algorithm, embedding and detection operations are both executed entirely in the compressed domain, with no need for the decompression process. The new criteria employing statistical invisibility of contiguous frames is used to adjust the embedding strategy and capacity, which increases the security of proposed algorithm. Therefore, the collusion resistant properties are obtained. Video steganalysis with closed loop feedback manner is design as a checker to find out obvious bugs. Experimental results showed this scheme can be applied on compressed video steganography with high security properties. 1. Introduction Steganography is the transmission of a secret message hidden within an ordinary carrier without revealing its existence. The container (cover file) may be a digital still image, audio file, or video file. Once the secret message has been embedded, it may be transferred across insecure lines or posted in public places. Usually, the data rate of covert data transmission using steganography is low in order to keep the covert data imperceptible within the cover medium. This data rate is somewhat proportional to the volume of the cover medium. For this reason, digital video is a convenient choice for steganography.

Nowadays, given the high degree of collaboration and cooperation in modern information systems such as emerging multimedia sensor networks, covert communications becomes a greater threat to forensic analysis than ever. It is imperative to investigate methods to detect and discourage covert communications such as steganography in multimedia networks that acquire highly correlated data. This paper will focus on the particular problem of the compressed video steganography.

General speaking, digital video appears in two main distinct encoding formats: the uncompressed and the compressed. The most popular compressed format by far is motion compensated compressed video, specifically the widely accepted standard MPEGx. It achieves compression through the elimination of temporal, spatial and statistical redundancies and with this compression operation. The video bit-stream consists of variable length codes (VLC) that represent various video segments.

For video stream usually being offered in compressed form, steganography algorithms that are not applicable in compressed bit-stream would require complete or at least partial decompression [1-4]. This is an unnecessary burden best avoided. If the requirement of strict compressed domain stegano-graphy is to be met, the steganography needs to be embedded in the compressed domain.

Nowadays, there are large amount of video watermarking algorithms been proposed. And some of them are applied for compressed video[5-9].

To be useful, a steganographic technique should not be easily detectable. If the existence of secret message can be detected with a probability higher than random guessing, the corresponding steganographic technique is considered to be invalid. Similar to cryptography, steganography may suffer from the attack method (steganalysis). Much of the research work in the field of steganalysis has been carried out on images. One approach is based solely on the first order statistics and is applicable only to idempotent embedding[10-13]. Another major stream is based on the concept of blind steganalysis, which is formed by blind classifiers [14-16]. The classifier should be trained to learn the differences between cover and stego-image features at first. There are also two video steganalysis methods have been proposed by Deepa et al.[17, 18] using collusion principle. And in ref [19] Deepa obtain some new video statistical invisibility properties, which inspired us to design this steganography.

In this paper, we propose a secure compressed video steganography architecture taking account of video

The Third International Conference on Availability, Reliability and Security

0-7695-3102-4/08 $25.00 2008 IEEEDOI 10.1109/ARES.2008.140

1384


0-7695-3102-4/08 $25.00 2008 IEEEDOI 10.1109/ARES.2008.140

1384


0-7695-3102-4/08 $25.00 2008 IEEEDOI 10.1109/ARES.2008.140

1382

Authorized licensed use limited to: Sun Microsystems- [email protected]. Downloaded on February 20,2010 at 04:35:49 EST from IEEE Xplore. Restrictions apply.

statistical invisibility. Also the architecture is with a steganalysis module, operated in a closed-loop manner to enhance the anti-steganalysis capability of the stego-video with data embedded.

This paper is organized as follows: Section 2 describes the framework of our secure steganography system. In Section 3, the embedding mechanism is described in detail. The Section 4 is the perform analysis. We give the experimental results in Section 5. In Section 6, a conclusion is drawn finally. 2. Framework for CVSS

parsersequence video

detector change Scene

embedder Message

U videoover C

M messageSecret

K eyK

issteganalys Video

Xvideo tegoS

Fig1. Framework for CVSS

As shown in the Fig. 1, the framework for CVSS consists of four function parts, the video sequence parser, the scene change detector, the secret message embedder and the video steganalysis.

Compressed video sequence achieves compression through the elimination of temporal, spatial and statistical redundancies with the use of motion compensation, block quantization inside a discrete cosine transform (DCT), and Huffman run-level encoding. With this compression, the video bit-stream consists of variable length codes (VLCs). The cover video sequence U pass through the sequence parser module at first, which has a VLD (variable length decoder), to be parsed as VLC code denotes the corresponding various segments of the video including, intracoded macroblocks, DCT coefficients and motion vectors, etc.

After that, the second module, scene change detector, can divide the sequence into several slow speed and single scene video sub-sequences. In the video sequence, a shot is defined as one object is shot with the same camera in a same place. The scene is

defined as several continuous shots. The pictures in one scene have the similar features. Scene change detector is to find out the scene change point among the group of pictures. In proposed framework, DC coefficients are used in the scene change detector module for the real time requirement. Embedding operations are taken place in I frames, therefore detector searching the scene change point among I frames in the compressed format.

The I frames in MPEG standard is coded in intra-frame manner, we can obtain the DC picture with abstracting the DC coefficients from the DCT coefficient codes. Eq1 describes the compare method between two conjoint I frames:

=

+

++

+

=Nk

ii

iiii kHkH

kHkHIIHD 1 21

21

1 ))()(())()((),( 1

where, iI , 1+iI means the ith and thi 1+ I frames, iH and 1+iH are histograms of DC pictures from the ith and thi 1+ I frames. If the ),( 1+ii IIHD is the peak value the two I frames are from different scenes, therefore the scene change point is found. Also the variances )var( iU of each DC picture from I frame will be calculated and provided to the embedding process module.

With the third module, message embedder, secret message M is hidden into the compressed video sequence without bringing perceptive distortion. There have two different embedding modes, inter-frame and intra-frame. To avoid the collusions attack[19], statistical invisibility is employed to adjust the capacity of the secret message adaptively. Also the chaotic dynamic system and VLC maps embedding algorithm will be used at the same time. This module is the key part in the framework, and will be described detailedly in the next section.

The last module, video steganalysis, is used as a checker for the message hiding security in the stego-video sequence. With the principle of collusion [17], deviation of the DC and AC coefficients between contiguous frames is statistical visible. Also the correlation between frames in one scene is used as another measurement of the data change. Using the first and higher order measurements statistics as input of the artificial intelligence classifier, the embedding of message can be detected in the stego-video sequence with a constant probability. If the stego-video cannot pass through the checker, the scale factor must be adjusted to manipulate the strength of hidden message.

With these four modules, we can obtain the final stego-video X with anti-steganalysis capability by holding the original statistical invisibility properties.

138513851383


3. VLC Domain Embedding Mechanism 3.1 Embedding Position Selection

In the VLC table in the MPEG standard, each codeword corresponds to a run-level pair, denoted as (run, level). During a video encoding stage, the pixel values in the spatial domain are 88 discrete cosine transformed (DCT). After quantization, the content of each 88 block is scanned in a zigzag manner. Each scanned non-zero integer has to be converted into a so-called (run, level) pair. Having the run-level pairs for all the non-zero quantized coefficients in an 88 block, one can consult the VLC table and convert them into a bitstream.

It is easy to see that if one chooses run-level pairs as potential hiding places, and then it will be feasible to modify the level value only. Based on the MPEG coding rule, if a run value is modulated, then this indicates that the number of preceding zeros has been altered. Basically, this kind of update will lead to serious consequences because the positions of all subsequent non-zero coefficients will be shifted. Under these circumstances, if these shifted coefficients are decoded back to the spatial domain, the resultant frame will be quite different from its original. However, if we update the level value, then only its magnitude will be changed.

Another issue regarding video steganography is the selection of a proper color channel for embedding. Usually, the frames of a video sequence will be split into Y, Cb and Cr channels in the MPEG coding stage. According to different color resampling rules, it is possible for the ratio Y:Cb:Cr be set to 4:2:2 or 4:2:0. Under these circumstances, the only unchanged channel is the Y channel. Thus, when embedding message, we prefer to exploit the Y channel as the host channel.

In addition to the above selections, choosing an appropriate frame type among I-frame, P-frame or B-frame for hiding message is also a crucial issue. Usually, a conventional video consists of a number of GOPs. Each GOP is composed of one I-frame and several B-frames and P-frames. A typical I-frame adopts intra coding, which means it does not refer to any other frames. Different from an I-frame, a P-frame only refers to its nearest preceding I- or P-frame. As for a B-frame, it refers to the nearest preceding and succeeding I or P-frame. In a conventional MPEG format, the content of a B- or P-frame is the so-called residual error between the current frame and the frame to which it refers. Therefore, only an I-frame can hold complete information. In this paper, we choose to

embed message into the I-frames of an MPEG compressed video sequence. 3.2 Analysis of Compressed Domain Embedding

In this subsection, we study the effect of compressed domain embedding strategy on the variance of the pixel values in spatial domain.

From the definition of DCT transform, the variance change caused by every DCT coefficients modification.

Let ),( yxf denotes the spatial pixel value of the yx, position in the video I picture. And ),( yxf

denotes the increment of the pixel value. )),(( yxfD denotes the variance of the whole I picture pixel values. And the )),(( yxfD , the increment of )),(( yxfD . With modification in 1, the change of variance of the whole I picture pixel values can be deduced as following:

(2) ]))2

)12(cos2

)12(cos),(((

)2

)12(cos2

)12(cos),(()[()(4

))2

)12(cos2

)12(cos),()()(2((

)2

)12(cos2

)12(cos),()()(2(

)),(()),((

)),(()),(),((

))),(()),(),(((

)),((),(

)),(),(()),(),((

)),((

2

222

21

1

1

1

21

1

1

1

22

22

22

22

22

Ny

MxFE

Ny

MxFEcc

MN

Ny

MxFcc

MNE

Ny

MxFcc

MNE

yxfEyxfE

yxfEyxfyxfEyxfEyxfyxfE

yxfEyxEfyxfyxfEyxfyxfE

yxfD

N N

N N

+

+

+

+=

+

+

+

+=

=++

+=+

++=

=

=

=

=

where, 1-M,1,2, x 1

0 x,2

1)(),(

=

=

=

"ycxc .

Where, the M and N is the size of frame picture. If one selected DCT coefficient is changed by 1, the change of variance in the spatial domain can be obtained from Eq (2).

Multiplying with the intra-code quatilized matrix, the effect on variance of each quantilized DCT coefficient can be obtained.

Table. 1. effect on variance of each quantilized DCT coefficient

u v 1 2 3 4

1 0 1.4142 1.5343 1.9445 2 1.4142 1.0000 1.3750 1.5000 3 1.5343 1.3750 1.6129 1.6875 4 1.9445 1.3750 1.6250 1.6875 5 1.8695 1.6250 1.6819 1.8125

138613861384


6 2.2981 1.6875 1.8125 2.0000 7 2.2872 1.6875 1.8118 2.1250 8 0.0000 0.0000 0.0000 0.0000 u v 5 6 7 8

1 2.2094 2.3865 2.5511 0.0000 2 1.6875 1.8125 2.1250 0.0000 3 1.8065 2.1250 2.1241 0.0000 4 1.8125 2.1250 2.3125 0.0000 5 1.9970 2.1875 2.4995 0.0000 6 2.1875 2.5000 3.0000 0.0000 7 2.3746 2.8750 3.4999 0.0000 8 0.0000 0.0000 0.0000 0.0000

Fig 2 is the 3D presentation of the data in Table 1.

It is can be concluded from the Fig, the effect types on variance of every DCT coefficient can be divided into four different groups. For the effect is linear and accumulatable. So the proposed embedding scheme manipulates the change magnitude of pixel values variance by adjusting the embedding payload is useful, which meet the secure principle that variance of message embedded should be adaptive according to the variance of the cover video frame

Fig 2 effect on variance of each DCT coefficient

3.3 Compressed Video Secure Steganography scheme In the proposed algorithm, the positive even and negative odd DCT coefficients represent zero and positive odd and negative even coefficients represent one. And the value 1 should be modified into -1 or 2, and value -1 should be modified into -2 and 1. Algorithm: Compressed video secure stego-system Input: Cover video U , key stream key , message S Output: Stego-video X , adjust coefficient . Preprocess: Step 1: Encrypt message

C keyE ( M ); Step 2: For each I frame iF calculate the pixel value

variance 2iF with the DCT DC coefficients;

Step 3: allot payload to each frame according to 2iF

;22

ii

F

Fi

Ss

Step 4: for each is , arrange the message bit to available DCT AC coefficients, and estimate the change of special pixel values variance iD ; Step 5: calculate the correlation of 2

iF and iD ;

if TDCorr iFi

])1(

)1(2)1(

214

)1(21[ 2

222

222

MM

z

U zzz

z

zzz

Fig.6 PSNR and correlation change magnitude of the

stego-video after message embedding 6. Conclusions In this paper, one new secure and file-size preserving compressed domain steganography is proposed. Embedding and detection are both done entirely in the compressed domain to meet the real time requirement. Change of the spatial pixel values variance can be estimated in the compressed domain, and the embedding payload is allotted according to the variance of each cover frame. Therefore the correlation of the continuous frames is unchanged. The performance of the steganography algorithm is studied and the experimental results showed this scheme can be applied on compressed videos with no noticeable degradation in visual quality.

References

[1] F Hartung, B. Girod Watermarking of uncompressed and compressed video. Signal Processing, Special Issue on Copyright Protection and Access Control for Multimedia Services, 1998, 66 (3): 283-301.

[2] G. C. Langelaar and R. L. Lagendijk. Optimal Differential Energy Watermarking of DCT Encoded Images and Video. IEEE Trans. on Image Processing, 2001, 10(1):148-158.

[3] B. Chen and G. W. Wornell, Quantization index modulation: a class of provably good methods for digital watermarking and information embedding. IEEE Transactions on Information Theory, 2001, 47(4):14231443.

[4] Y. J. Dai, L. H. Zhang and Y. X. Yang. A New Method of MPEG Video Watermarking Technology. International Conference on Communication Technology Proceedings (ICCT), 2003.

[5] C. S. Lu, J. R. Chen and K. C. Fan. Real-time frame-dependent video watermarking in VLC domain. Signal Processing: Image Communication 20 (2005) 624-642.

[6] G. C. Langelaar. Watermarking Digital Image and Video Data. IEEE Signal Processing Magazine, 2000, (17)5:20-46.

[7] B. G. Mobasseri and M. P. Marcinak.Watermarking of MPEG-2 Video in Compressed Domain Using VLC Mapping. ACM Multimedia and Security Workshop 2005, New York, NY, August 2005.

[8] H. H. Liu, L. W. Chang. Realtime digital video watermarking for digital rights management via modification of VLCs. Proceedings of the 11th International Conference on Parallel and Distributed Systems (ICPADS'05), 2005.

[9] D. P. Ye, C. F. Zou, Y. W. Dai, Z. Q. Wang. New adaptive watermarking for real-time MPEG videos. Applied Mathematics and Computation 185(2007): 907-918.

[10] A. Westfeld, A. Pfitzmann. Attacks on Steganographic Systems, Lecture Notes in Computer Science, vol. 1768, Springer-Verlag, Berlin, 2000, pp. 61-75.

[11] J. Fridrich, R.Du, L. Meng. Steganalysis of LSB Encoding in Color Images, Proceedings IEEE International Conference on Multimedia and Expo, July 30-Auguest 2, 2000, New York City, NY.

[12] J. Fridrich, M. Goljan. Practical Steganalysis of Digital Images State of the Art, In Delp III, Security and Watermarking of Multimedia Contents IV. Volume 4675 of Proc. SPIE. (2002) 1-13.

[13] S. Dumitrescu, X. L. Wu, Z Wang. Detection of LSB Steganography via Sample Pair Analysis. LCNS, vol.2578, Springer-Verlag, New York, 2002:355-372.

[14] H. Farid, L. Siwei. Detecting Hidden Messages Using Higher-Order Statistics and Support Vector Machines. In: Petitcolas, F.A.P. (ed.): Information Hiding. 5th International Workshop. Lecture Notes in Computer Science, Vol. 2578. Springer-Verlag, Berlin Heidelberg New York (2002) 340-354

[15] J. Harmsen, W. Pearlman. Steganalysis of additive noise modelable information hiding, in Proc. SPIE Security and Watermarking of Multimedia Contents V, Santa Clara, CA, Jan. 2003, vol. 5022

[16] I. Avcibas, B. Sankur, K. Sayood. Statistical evaluation of image quality measures, J. Electron. Imaging, vol. 11, no. 2, pp. 206223, Apr. 2002.

[17] U. Budhia, D. Kundur, T. Zourntos. Digital Video Steganalysis Exploiting Statistical Visibility in the Temporal Domain. IEEE Trans. Information Forensics and security, 2006, (1)4:502-516.

[18] J. S. Jainsky, D. Kundur and D. R. Halverson. Towards Digital Video Steganalysis using Asymptotic Memoryless Detection. MM&Sec07, September 2021, 2007, Dallas, Texas, USA.

[19] K. Su, D. Kundur, and D. Hatzinakos, Statistical invisibility for collusion-resistant digital video watermarking, IEEE Trans. Multimedia, 2005, (7)1:43-51

138913891387


Documents

Secure Steganography in Compressed Video Bitstreams