A New Video Steganography Algorithm Based on the Multiple Object Tracking and Hamming Codes Ramadhan J. Mstafa, IEEE Student Member
Department of Computer Science and Engineering University of Bridgeport Bridgeport, CT 06604, USA
[email protected] Abstract—In the modern world, video steganography has become a popular option for secret data communication. The performance of any steganography algorithm is based on the embedding efficiency, embedding payload, and robustness against attackers. In this paper, we propose a new video steganography algorithm based on the multiple object tracking algorithm and Hamming codes. The proposed algorithm includes four different stages. First, the secret message is preprocessed, and Hamming codes (n, k) are applied in order to produce an encoded message. Second, a motion-based multiple object tracking algorithm is applied on cover videos in order to identify the regions of interest of the moving objects. Third, the process of embedding 3 and 6 bits of the encoded message into the 1 LSB and 2 LSBs of RGB pixel components is performed for all motion regions in the video using the foreground mask. Fourth, the process of extracting the secret message from the 1 LSB and 2 LSBs for each RGB component of all moving regions is accomplished. Experimental results of the proposed video steganography algorithm have demonstrated a high embedding efficiency and a high embedding payload. Keywords- Video steganography; Motion-based multiple object tracking; Hamming codes; Embedding efficiency; Eembedding payload
I. INTRODUCTION Recently, people communicate over the Internet and share private information. This secret information should be protected through a secure technique that blocks the data from intruders and hackers. Steganography is a technique that protects any secret message from an unintended recipient’s suspicion within any data form. However, many steganalytical detectors have been invented that detect a secret message from an unsecure steganography algorithm. In order to avoid the secret data from being detected by steganalytical tools, the steganography algorithm must be efficient. Every successful steganography algorithm should contain an embedding efficiency, an embedding payload, and robustness in order to work against attackers [1, 2]. The steganography algorithm containing a high embedding efficiency will reduce a hacker’s suspicion of finding the hidden data, and will be difficult to detect through steganalysis detectors. In addition, accurate visual quality of the stego data and a low modification rate of the cover data will improve the embedding efficiency [3]. The embedding efficiency includes: perceptual quality,
Khaled M. Elleithy, IEEE Senior Member
Department of Computer Science and Engineering University of Bridgeport Bridgeport, CT 06604, USA
[email protected] complexity, and security. The embedding efficiency is directly affected by the security of the steganographic algorithm. However, if any obvious distortion of the cover data after the embedding process occurs, then the result may be an increase in the attention of the hackers [4]. Another component of a successful steganography algorithm is a high embedding payload. The embedding payload is defined as the amount of secret information that is required to be embedded inside the cover data. Furthermore, an algorithm that contains a high embedding payload will have an extensive capacity to hide a secret message. In traditional steganographic algorithms, an embedding efficiency and an embedding payload are opposites. Increasing the capacity of the secret message will decrease the visual quality of stego videos resulting in a weakened embedding efficiency. Both factors should be considered. The deciding factors depend on the steganography algorithm and the user’s requirements [4, 5]. In order to increase the embedding payload with the low modification rate of the cover data, many steganography algorithms have been proposed using alternative methods. These algorithms use block codes and matrix encoding principles such as BCH codes, Hamming codes, cyclic codes, Reed-Solomon codes, and Reed-Muller codes [6]. The contribution of this paper introduces a new video steganography algorithm based on the multiple object tracking algorithm and Hamming codes. Our proposed steganography algorithm offers a reasonable trade-off between the perceptual quality and an embedding payload. The remainder of this paper is organized as follows: Section 2 presents some related works. Section 3 explains Hamming codes. Section 4 reviews motion-based multiple object tracking algorithm. Section 5 presents the proposed steganography algorithm. Section 6 illustrates and explains the experimental results. Section 7 is the conclusion. II. RELATED WORK Cheddad et al. proposed a skin tone video steganography algorithm based on the YCbCr color space. YCbCr color space is a useful color transformation, which is used in many techniques such as compression and object detection methods. The correlation between three color channels (RGB) is removed, so that the intensity (Y) will be separated from colors chrominance blue and red (Cb and Cr). After the human skin regions are detected, the only Cr of these regions will be utilized for embedding the secret message. [7]. Overall, the algorithm has a low embedding payload because
it has embedded the secret message innto the only Cr component of the skin region. Khupse et al. proposed an adaptive video steganography scheme usiing steganoflage. The steganography scheme has been useed in region of interest video frames. Khupse et al. used hum man skin color as a cover data for embedding the secrett message. The morphological dilation and filling operatioon methods have been used as a skin detector. After viddeo frames have converted to YCbCr color space, the fraame that has the minimum mean square error will be seelected for data embedding process. Only the Cb com mponent of this particular frame will be picked for embeedding the secret message [8] . This scheme is very lim mited in capacity because only one frame is selected for thee data embedding process. Zhang et al. proposed an efficientt embedder using BCH codes for steganography. The embeddder conceals the secret message into a block of cover data.. The embedding process is completed by changing various ccoefficients in the input block in order to make the syndrome values null. The efficient embedder improves both storagge capacity and computational time compared with otther algorithms. According to the system complexity, Zhhang’s algorithm improves the system complexity from expoonential to linear [9]. There is flexibility for both embeddinng efficiency and embedding payload in the previously mentiioned algorithms. This flexibility can be used by our propoosed algorithm to improve the algorithm’s performance even ffurther.
channel, a message M, which inccludes of 11-bit, will be multiplied by the generator matrix G, G and then, manipulated by having modulo of 2. The co odeword X of 15-bit is obtained and ready to be sent. 1 At the receiver channel, the en ncoded data (message + parity) which is a codeword R of 15 5-bit will be received and checked for errors. Once the reeceived codeword R is multiplied by the parity-check mattrix H, modulo of 2 will then be applied. A syndrome vector Z , , , of 4-bit is obtained. If the received message is correct, then Z must have all zero bits (0000); otherwise,, during the transmission, one or more bits of the received message m might be flipped. In that case, the error correction pro ocess must occur. 2
Where
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
1 1 0 0
0 1 1 0
0 0 1 1
1 1 0 1
1 0 1 0
0 1 0 1
1 1 1 0
0 1 1 1
1 1 1 1
1 0 1 1
1 0 0 1
The reason of using parity bits in i the Hamming codes is to protect the message during comm munication. In Hamming codes (15, 11), 7 bits of the messaage are used to calculate each of parity bit, which is illustrateed in the Fig. 1.
III. HAMMING CODES In this section, the Hamming codes teechnique will be explained and discussed through a specificc Hamming code (15, 11). Hamming codes are defined as one of the most powerful binary linear codes. These typees of codes can detect and correct errors that occur in the binary block of data during the communication between pparties [10]. The codeword includes both original and exxtra data with a minimum amount of data redundancy, andd is the result of the encoded message that uses the Hamming code technique. In general, if p is parity bits of a positive integer number 2 ; then, the length of the ccodeword is 1. The size of the message that needs to be encoded is 2 1. The number oof parity bits that defined as 2 must be added to the message is with the rate of / [4, 11]. In this paper, Hamming codes (15, 11) are used (n=15, k=11, and p=4), which can correct the iddentification of a single bit error. A message of size , ,…, is encoded by adding , , , extraa bits as parity to become a codeword of 15-bit length. T The codeword is prepared to transmit through a communiccation channel to the receiver end. The common combinationn of both message and parity data using these type of codess is to place the parity bits at the position of 2 (i=0, 1, ..., nn-k) as follows: , , , , , , , , , , , , , , During the encoding and decoding processes, the generator matrix G and parity-check mattrix H are being used by Hamming codes (15, 11). At the transmitter
Figure 1. Venn diagram of the Ham mming codes (15, 11).
IV.
MOTION-BASED MULTIPL LE OBJECT TRACKING
Due to its various applications, computer c vision is one of the fastest emerging fields in computer science. The detection and tracking of moviing objects within the computer vision field that has reccently gained significant attention. The tracking of movin ng objects is commonly divided into two major phases: 1) detection of moving me, and 2) association of objects in an individual video fram these detected objects throughout all a video frames in order to construct complete tracks [12].
In the first phase, the background subtraction method, based on the Gaussian Mixture Model (GMM), is used to detect the moving objects. GMM is a probability of density function equal to a weighted sum of component Gaussian densities. The background subtraction method computes the differences between consecutive frames that generate the foreground mask. Then, the noises will be eliminated from the foreground mask by using morphological operations. As a result, the corresponding moving objects are detected from groups of connected pixels. The second phase is called data association. It is based on the motion of the detected object. A Kalman filter is utilized to estimate the motion of each trajectory. In each video frame, the location of each trajectory is predicted by the Kalman filter. In addition, the Kalman filter is used to determine the probability of a specific detection that belongs to each trajectory [12]. V. THE PROPOSED STEGANOGRAPHY ALGORITHM In this section, we present a new video steganography algorithm based on the multiple object tracking algorithm and Hamming codes (15, 11). Our proposed steganography is divided into the following four stages: A. Secret Message Preprocessing Stage In this work, a large size text file is used as a secret message, and it is preprocessed before the embedding stage. Here, the whole characters in the text file are converted into ASCII codes in order to generate an array of binary bits. Then, for security purposes, the binary array is encrypted by using a key (Key1) that represents the size of the secret message. This process will encode the message and protect it from attackers. Since the binary linear block of Hamming codes (15, 11) are used, the encrypted array is divided into 11-bit blocks. Then, every block is encoded by the Hamming codes (15, 11) that will produce 15-bit blocks. The size of the message is extended by adding four parity bits into each block. Another key (Key2) is utilized to generate randomized 15-bit numbers, and each number is XORed with the 15-bit encoded block. The security of the proposed algorithm will be improved by using two keys, Hamming codes, and XOR operation. B. Motion-Based Multiple Object Tracking Stage The motion-based multiple object tracking algorithm has been previously explained in Section 4. The process of identifying the moving objects in the video frames must be performed when motion object regions are used as cover data. This process is achieved by detecting each moving object within an individual frame, and then associating these detections throughout all of the video frames. The background subtraction method is applied to detect the moving objects. Then, the Kalman filter is used to predict estimation trajectory of each moving object. C. Data Embedding Stage In each video frame, the cover data of the proposed algorithm is the motion objects that are considered as regions
of interest. The motion regions are identified through the video frames after they are detected and tracked. The region of interest changes in every frame based on the size and the number of the moving objects. The motion-based multiple object tracking algorithm is applied in order to predict trajectories of all moving objects. In each video frame, the background subtraction method is administered to generate a foreground mask which will determine the regions of the moving objects. Then, the R, G, and B components of each motion object’s pixels are used for embedding purposes. In the proposed algorithm, the 1 LSB and 2 LSBs are utilized in order to embed 3 and 6 bits of the secret message in each motion pixel. Moreover, in order to transmit keys to the receiver party, both keys are embedded into the non-motion region of the first video frame. Upon completion, the stego frames will be reconstructed in order to produce the stego video format that transmits via the communication medium to the receiver party. Fig. 2 shows the block diagram of the data embedding stage. D. Data Extraction Stage The process of the data extracting stage is illustrated in Fig. 3. In order to retrieve a secret message correctly, the stego video is divided into frames through the receiver, and then two keys are extracted from the non-motion region of the first frame. To predict trajectories of motion objects, the motion-based multiple object tracking algorithm is applied again by the receiver. Moreover, in each video frame, a foreground mask that is similar to the embedding stage’s mask is produced by using the background subtraction method. Then, the process of extracting the hidden message is conducted by taking out 3 and 6 bits from the 1 LSB and 2 LSBs of RGB color components in each motion object’s pixels of all the video frames. The extracted bits from all the video frames are stored in a binary array. The binary array is divided into 15-bit blocks. Each block will be XORed with the 15-bit number randomly generated by Key2. The results of the 15-bit blocks are decoded by using the Hamming (15, 11) decoder to produce 11-bit blocks. Since the sender has encrypted the secret message, the obtained array is decrypted by using Key1. The final array is divided into an 8-bit code (ASCII) in order to generate the right characters of the original message. VI. EXPERIMENTAL RESULTS AND ANALYSIS Three S2L1 video sequences of different views (View1, View3, and View4) were used from the well-known PETS2009 dataset [13]. The implemented videos contain moving objects which are taken by different stationary cameras. Experimental results are obtained by using the R2013a version of the MATLAB software program. The videos contain a 768x576 pixel resolution at 30 frames per second, and a data rate of 12684 kbps. Each cover video sequence contains 795 frames. In all the video frames, the secret message appears as a large text file split in accordance with the size and number of the moving objects.
Figure 2. Block diagram of the data embedding stage of the proposed algorithm.
Figure 3. Block diagram of the data extraction stage of the proposed algorithm.
A. Visual Quality The visual quality of the proposed algorrithm is measured by applying the Peak Signal to Noise Ratioo (PSNR) metric. PSNR is a non-perceptual objective metric tthat measures the difference between the original and the distoorted videos. It is calculated as follows: 3
10
And Mean Square Error (MSE) is calculaated as follows: ∑
∑
∑
, ,
, ,
4
C and S refer to the cover frame aand stego frame, respectively. In addition, m and n are ddefined as video resolutions, and h indicates the R, G, and B color channels (k=1, 2, and 3).
Figure 5. The PSNR comparison of the View3 video.
Fig. 4 shows the PSNR comparison oof the first video (View1) when using 1 LSB and 2 LSBs of each motion object’s RGB pixels. Here, the PSNR values equal 47.73 dB for 1 LSB and 40.45 dB for 2 LSBs. Fig. 5 illustrates the PSNR comparisoon of the View3 experiment when using 1 LSB and 2 LSBss of each motion pixel in the video frames. The PSNR valuess equal 50.93 and 43.88 dBs for 1 LSB and 2 LSBs, respectiveely. Fig. 6 shows the PSNR comparison of the View4 video when using 1 LSB and 2 LSBs of each motiion object’s RGB pixels. Here, the PSNR values equal 51.35 ddB for 1 LSB and 44.16 dB for 2 LSBs. The third experiment (View4) has bettter visual quality among other experiments because it has ffewer regions of moving objects than others. This means thatt View4 video can embed less size of the secret data thann the other two experiments. Overall, the stego videos’ vissual qualities are close to the original videos’ visual qualitiess due to the high values of PNSRs for our proposed algorithm m.
Figure 6. The PSNR comparison of the View4 experiment.
B. Embedding Payload According to the reference [14], our proposed algorithm has a high embedding payload. Heree, the average of obtained hiding ratios for three experiments is i 3.37%. The size of the hidden secret message in each Viiew1, View3, and View4 videos using 1 LSB is 31.38, 14.6 62, and 12.95 Megabits, respectively. Moreover, when using g 2 LSBs, the amount of the secret message in each View1, View3, and View4 5, and 25.92 Megabits, experiments will be 62.77, 29.25 respectively. The hiding ratio can bee calculated as follows: 100%
5
Fig. 7, 8, and 9 illustrate the daata embedding payload of the proposed steganography algo orithm for each View1, View3, and View4 experiments. These T three figures have shown the comparison of the emb bedding capacity of each video when 1 LSB and 2 LSBs of th he moving objects’ pixels Figure 4. The PSNR comparison of the View1 experiment.
are utilized. The 2 LSBs were implemennted in order to double the amount of the secret message in eeach experiment.
Figure 7. The embedding payload comparison of the View1 experiment.
VII. CONCLUSIO ON This paper introduces a new w video steganography algorithm based on the multiple object o tracking algorithm and Hamming codes. The four algo orithm stages include: 1) the secret message preprocessing sttage; 2) the motion-based multiple object tracking algorith hm stage; 3) the data embedding process stage; and 4) the t data extracting stage. The performance of the proposeed algorithm is verified through a series of experiments. Our O experimental results demonstrated that the proposed algorithm has a high h values of obtained embedding efficiency based on high PSNRs. Furthermore, the prop posed algorithm used protection methods on the secrett message prior to the embedding process creating additional efficiency. m has a high embedding Moreover, the proposed algorithm payload due to the sizable amou unt of the hidden secret message. In our future work, we will apply the proposed n in order to verify the algorithm in the frequency domain robustness of the algorithm againstt various attacks such as video processing attacks and artificiial noises. REFERENC CES
Figure 8. The embedding payload comparison of tthe View3 video.
Figure 9. The embedding payload comparison of the View4 experiment.
[1] R. J. Mstafa and K. M. Elleithy, "A novel video steganography algorithm in the wavelet domain based on the KLT traccking algorithm and BCH codes," in Systems, Applications and Technology Conference (LISAT), 2015 IEEE Long Island, 2015, pp. 1-7. [2] R. J. Mstafa and K. M. Elleithy, "An Efficient Video Steganography merican Society for Engineering Algorithm Based on BCH Codes," in Am Education (ASEE Zone 1), 2015 Zone 1 Con nference, Boston, 2015, p. 10. [3] H. Yuh-Ming and J. Pei-Wun, "Two imp proved data hiding schemes," in Image and Signal Processing (CISP), 2011 4th International Congress on, 2011, pp. 1784-1787. [4] R. J. Mstafa and K. M. Elleithy, "A highly secure video steganography using ations and Technology Conference Hamming code (7, 4)," in Systems, Applica (LISAT), 2014 IEEE Long Island, 2014, pp.. 1-6. [5] R. J. Mstafa and C. Bach, "Informaation Hiding in Images Using Steganography Techniques," in American Society S for Engineering Education Northeast Section Conference, 2013. gh payload video steganography [6] R. J. Mstafa and K. M. Elleithy, "A hig algorithm in DWT domain based on BC CH codes (15, 11)," in Wireless Telecommunications Symposium (WTS), 20 015, 2015, pp. 1-8. [7] A. Cheddad, J. Condell, K. Curran, and P. McKevitt, "Skin tone based g the YCbCr colour space," in Steganography in video files exploiting Multimedia and Expo, 2008 IEEE Interna ational Conference on, 2008, pp. 905-908. [8] S. Khupse and N. N. Patil, "An adaptive steeganography technique for videos using Steganoflage," in Issues and Challlenges in Intelligent Computing Techniques (ICICT), 2014 International Co onference on, 2014, pp. 811-815. [9] Z. Rongyue, V. Sachnev, M. B. Botnan, K. K Hyoung Joong, and H. Jun, "An Efficient Embedder for BCH Coding for f Steganography," Information Theory, IEEE Transactions on, vol. 58, pp. 7272-7279, 2012. unath, "Matrix Embedding With [10] A. Sarkar, U. Madhow, and B. S. Manju Pseudorandom Coefficient Selection and Error Correction for Robust and F and Security, IEEE Secure Steganography," Information Forensics Transactions on, vol. 5, pp. 225-239, 2010. [11] C. Chin-Chen, T. D. Kieu, and C. Yung-Chen, "A High Payload Steganographic Scheme Based on (7, 4) Haamming Code for Digital Images," in Electronic Commerce and Security, 20 008 International Symposium on, 2008, pp. 16-21. Object tracking: A survey," Acm [12] A. Yilmaz, O. Javed, and M. Shah, "O computing surveys (CSUR), vol. 38, p. 13, 2006. 2 [13] J. Ferryman, in Pets 2009 dataset: Perform mance and evaluation of tracking and surveillance, 2009. [14] L. Tse-Hua and A. H. Tewfik, "A noveel high-capacity data-embedding system," Image Processing, IEEE Transacctions on, vol. 15, pp. 2431-2440, 2006.
