A multiplicative video watermarking robust to H.264/AVC compression standard
Introduction
Achieving higher video quality at the same bit-rate has resulted in overgrowing application of the H.264 video compression standard. For this innovative development, sophisticated algorithms are required for H.264 authentication and copyright protection purposes. In this direction, many video watermarking algorithms have been proposed in the literature [[1], [2], [3], [4], [5], [6]].
Video watermarking algorithms can be divided into three groups including pixel domain, transformed domain, and compressed domain based on the place in which the embedding is directly applied. Pixel domain algorithms directly embed the watermark into video frame pixels. The main advantage of these algorithms is that they have low computational complexity and as a consequence can be implemented easily. However, they are less robust against attacks; in this regard, statistical and geometrical techniques are utilized to improve the robustness [[7], [8]]. In the transformed domain algorithms, the original video firstly undergoes a transform such as discrete Fourier transform (DFT), discrete cosine transform (DCT) or discrete wavelet transform (DWT). Then the watermark is embedded into the transformed coefficients [[9], [10], [11], [12], [13], [14]]. The watermarking in the compressed video is performed based on two different approaches: embedding during compression [[15], [16], [17], [18], [19], [20]], and embedding in the compressed video [21]. The interested readers are referred to [[22], [23], [24]] for a detailed survey on video watermarking in the compressed domain.
Despite video watermarking in the compressed domain, which may cause error propagation due to embedding, watermarking in the uncompressed domain is easier. In this case, the designer does not care about the video bit-rate. If the algorithm is implemented properly, it can be robust to various compression attacks. For embedding in the raw video, one can assume each frame as an independent still image. In this case, the algorithm does not provide great robustness against temporal attacks which may cause the watermark be vanished after compression. Therefore, watermarking is usually performed taking the advantage of temporal correlation between frames. In this manner, the watermark is embedded within a few frames which makes it more robust against compression attacks. A watermarking method for the raw video is proposed in [10] and [11]. In both methods, the original video is decomposed into groups of frames and each frame is divided into non-overlapping blocks. Applying three dimensional (3D) DCT transform and the quantization index modulation (QIM) method [25], the watermark is embedded in [11]. Since the watermark is embedded by manipulation of DC coefficients, this algorithm shows high robustness but low invisibility.
On the other hand, watermark embedding can be content dependent or independent. In the latter form, data hiding is performed using a mathematical operator which combines the watermark with a sequence of host features [26]. A popular example of this form is additive embedding where the watermark is spread via a pseudo-random (PN) sequence and then added to a sequence of consecutive host samples with a constant gain factor. Assuming the host samples to take Gaussian distribution and the attack is limited to additive white Gaussian noise (AWGN), the correlation decoder is optimal for watermark extraction [27]. In the former case, as the increase of watermark power causes more resistance against attacks, it is rational to match characteristics of the watermark to those of video features. A commonplace approach in this way is the multiplicative embedding where the energy of watermark samples is proportional to those of video features. The multiplicative embedding algorithms enjoy greater robustness against noise and gain attacks, compared to additive embedding methods [28]. According to [[29], [30], [31], [32], [33]], multiplicative methods are preferred to additive ones since the distortion proportional to the signal strength is more difficult to perceive. Thus multiplicative algorithms lead to more powerful watermark embedding while keep the quality of the watermarked signal at an acceptable level. Moreover, multiplicative methods are more secure than additive ones because the watermark of is host signal dependent. Consequently, it is more difficult to estimate the watermark by averaging on a set of watermarked signals. This attack is called the collision attack [34], and happened when copies of multiple works with the same watermark is available. Furthermore, since the power of watermark is proportional to the signal strength, greater coefficients tolerate more watermark and thus more power of watermark is inserted within the host signal. This is why multiplicative methods have greater robustness against most of attacks than the additive methods. Moreover, the multiplicative embedding rule automatically implements a simple contrast masking of Watsons perceptual model.
The maximum likelihood (ML) detection is prevalent for watermark extraction in multiplicative methods [27]. Two multiplicative watermarking methods are proposed in [9] and [35], where ML detection and side information are used for watermark extraction. A blind multiplicative watermarking method with ML detector is proposed in [36] that requires no side information for the sake of watermark extraction. In [10], the multiplicative watermarking rule in the LL subband of the wavelet transform is applied for data embedding. Both data embedding and extraction phases are designed based on the principal component analysis (PCA).
In this paper, a robust watermarking method for raw video signals has been proposed. Although video signals are often stored and transmitted in a compressed format, there are also applications in which video signals are stored in a raw format. Furthermore, there exists a need to provide security for uncompressed video before broadcasting and publishing. A trivial example is archiving procedures and releasing movies in cinemas. In fact, in the world broadcasting, videos are archived in a raw format. Once it is needed to broadcast a video, the streamer compresses the video signal with different bit-rates. Then, according to the available bandwidth suitable stream is broadcast. In order to preserve the copyright protection, watermarking should be applied on the raw video signal. The original video is decomposed into the groups of frames (GOP). Each frame is segmented to non-overlapping blocks. Every block forms a cube with other blocks in the same location and same GOP. A three-level 2D-DWT is applied to each cube where a single watermark bit is embedded into HH subband coefficients of each cube. We choose these position in order to guarantee the invisibility, while each watermark bit is spread over numerous blocks from different frames with the help of multiplicative embedding rule to address the robustness requirements [[35], [36]]. The ML detector applied to extract the watermark requires the distribution of the original signal. It has been shown that the wavelet detail coefficients can be well approximated by the generalized Gaussian distribution (GGD) model [[37], [38]]. A GGD model with the parameter (Laplacian distribution) for detail DWT coefficients is employed in this work. In order to satisfy the blindness requirement, i.e. estimating the parameters that our ML detector needs, the host signal is divided into two patches. The first patch hosts the watermark, while the other remains intact. The system parameters needed for extraction are estimated from the second patch of the signal. One of the innovations of the paper is that the decision of ML detector is based on the ratio of the sum of watermarked samples to the sum of intact samples. The use of this ratio makes the proposed method more robust against the AWGN and gain attacks. The method is compared with one of the state of the art methods [10] under different benchmarks and conditions. The proposed method shows superior performance against common attacks and also offer lower computational complexity in comparison with [10].
The rest of the paper is organized as follows. The statistical modeling of the watermarking system is presented in Section 2. The watermark embedding and the extraction procedure based on ML detector are explained in Section 3. Section 4 is dedicated to the analysis of the performance and efficiency of the proposed method. The experimental results are given in Section 5, and Section 6 concludes the paper.
Section snippets
System modeling
The system modeling of the proposed watermarking method is introduced in this Section. First, the distributions of the watermarking variables are calculated. Suppose that the host signal (wavelet HH coefficients here) has the identical and independent distribution (i.i.d.) of GGD with , or the Laplacian distribution. This claim will be confirmed later through experiments.
Consider as a zero mean Laplacian sequence with parameter . This sequence is divided into two patches based on the
Proposed method
A multiplicative watermarking algorithm for a raw video is proposed in this Section. The watermark bits are embedded into the HH3 coefficients of each cube. This design is proposed to address two main requirements of watermarking systems: invisibility and robustness. Selecting the HH subband coefficients guarantees the invisibility of the proposed method [39], while the robustness requirement is satisfied through multiplicative embedding in the host signal as will be explained.
Performance analysis and evaluation
The theoretical performance of the proposed watermark embedding and extraction framework is studied in this Section. First, a closed-form expression for the error probability of the ML detector is derived. Then, the expected performance of the proposed algorithm against AWGN is examined theoretically. Finally, the parameter selection analysis is given at the end of this Section.
Simulation results
In this section, the results of the implementation of the proposed method on the video sequences coastguard, bus, foreman, and mobile are presented. The sequences were YUV 4:2:2 with the CIF resolution (352 288) and the frame rate was 30 frames per second (fps). The watermark was embedded only into the luminance coefficients. The Haar wavelet filter was applied to find the 2D-DWT coefficients. Three-level 2D-DWT was applied on the 32 32 blocks and the watermark was embedded into the HH3
Conclusion
In this paper, a blind video watermarking method was proposed. With this method, the watermark is embedded in the HH subband coefficients of the three-level 2D-DWT of the video frames based on a multiplicative rule. The multiplicative rule is employed to better match the host characteristics to the watermark sequence resulting in more transparency. The statistical modeling of the host signal is applied for watermark extraction where ML detector is implemented. Embedding the watermark only into
References (42)
- et al.
Robust video watermarking of H.264/AVC
IEEE Trans. Circuits Syst. II
(2007) - et al.
Twenty years of digital audio watermarking — a comprehensive review
Signal Process.
(2016) - et al.
A new detector for contourlet domain multiplicative image watermarking using bessel k form distribution
J. Vis. Commun. Image Represent.
(2016) - et al.
A new multiplicative watermark detector in the contourlet domain using t location-scale distribution
Pattern Recognit.
(2018) - et al.
Digital Watermarking: from Concepts to Real-Time Video Applications
(1999) - et al.
Non-blind structure-preserving substitution watermarking of H.264/CAVLC inter-frames
IEEE Trans. Multimed.
(2014) - et al.
Digital cinema watermarking for estimating the position of the pirate
IEEE Trans. Multimed.
(2010) - et al.
A low complexity video watermarking in H.264 compressed domain
IEEE Trans. Inf. Forensics Secur.
(2010) - N. Zarmehi, M. Banagar, M. Akhaee, Optimum decoder for an additive video watermarking with Laplacian noise in H.264,...
- et al.
Everything you want to know about watermarking: from paper marks to hardware protection: from paper marks to hardware protection
IEEE Consum. Electron. Mag.
(2017)