Similar to information hiding in still images, luma DCT coefficients are usually used for information hiding by means of bit-plane replacement.
Ma et al. [77] propose to embed information in DCT coefficients quantized in I-frames (intensity). Based on the analysis of the relationship between the DCT coefficients and the distortion produced by the pixel values, several coefficients are paired to achieve the purpose of data embedding and distortion adjustment. The results show that the method is able to eliminate i-frame distortion drift, achieve higher payload, and reduce visual distortion. Lin et al. [78], as an extension of Ma's work, propose to embed two bits in the luma channel of selected macroblocks. The prediction mode (i.e. I4MB) and selected pixels in that macroblock are defined in their proposed mapping rules to achieve higher payload while maintaining video quality.
Earlier, Huang et al. [79] proposed an algorithm based on communication theory. It embeds information in the direct current DC coefficients followed by the low frequency alternating current AC coefficients. A similar technique was also proposed by Barni et al. [80], who defined video content as video object planes in the video object layer. The technique of Barni et al. computes a frequency mask, selects a pair of DCT coefficients, and divides them into two parts. For non-zero DCT coefficient parts, information is inserted into coefficients larger than a predefined threshold level. For the zero DCT coefficient part, the quantization parameters are manipulated to represent a hidden message. In the technique of Chung et al. [14], a histogram translation is applied to the DCT coefficients in the i frame, and the motion vectors in the adjacent macroblocks in the P/B frame are modified to achieve error concealment. Similarly, Shahid et al. [81] proposed to embed information by manipulating non-zero DCT coefficients within and between frames using different quantization parameters. In [82], Chen et al. [83] utilize Watson's visual mask construction and the payload estimation method of Lin et al. [78] to implement information hiding using selected DCT coefficients in frame i.
Thiesse et al. [84]–[86] hide the motion vector contention index (MVComp) in the chrominance and luma DCT coefficients to reduce the overall bitrate in H.264 video streams. A mapping rule based on the sum of DCT coefficients Si is proposed to control the change of bit rate and minimize the distortion by reducing the accuracy of motion estimation. MVComp is denoted by the parity of Si (the sum of DCT coefficients), and the prediction index Ii ∈ {0,1} is represented by adding hi to Si (when necessary):
The results show that the method achieves a good trade-off between luma texture prediction error propagation and visual quality in terms of chroma.
Meuel et al. [87] studied a similar technique to hide region of interest (ROI) information into quantized DCT coefficients. The ROI information is used to represent the important objects in the still image, and the ROI information is constructed based on the skin pixels (the boundary of the object in the still image).
Where Pu and Pv are Cb and Cr (chroma) components respectively, Pu and Pv are reference components, and d is a threshold that determines whether the current pixel is marked as a skin pixel. Its position, width and height values are embedded in the two non-zero DCT coefficient LSBs of the current frame. This technique achieves lossless reconstruction, but it turns out that the frame load is not enough to carry the entire region of interest information.
Similarly, Yin et al. [88] proposed a multi-directional interpolation technique using edge detection and residual information to hide information in edge pixels. This technique is designed for error concealment applications in still image decoders. Yilmaz et al. [89] proposed to hide quantized edge information (derived from neighboring macroblocks) for the purpose of error concealment. Based on [88] and [89], Kang et al. [90] used the parity embedding method in DCT coefficients to embed important information of macroblocks, including coding mode, reference frame, motion vector, etc., into the next frame. Li et al. [91] embed information in DWT coefficients for video watermarking. The scale coefficients of wavelet transform are used to embed low-resolution video frames, and the watermark information is embedded using wavelet coefficients. Besides, Wu et al. [92], [93] proposed the architecture, design and implementation of information hiding in the domain of still image and video. They recursively embed information in each video frame to handle uneven loads by selectively using modulation and multiplexing techniques in different regions.
Nakajima et al. [94] did not modify the non-zero DCT coefficients, but used the (zero) running components of the non-zero coefficients to embed information in the compressed video. For each block, calculate the position of the last non-zero coefficient (with respect to the Zigzag scan order), denoted by l. The value γ = log2(64−l) is the number of bits that can be embedded in the current block. The information is embedded by introducing a non-zero coefficient θ. θ is located at l+λ10, where λ10 is the decimal representation of γ. The sign and magnitude of θ can also be used for information hiding purposes.
For references marked in the text, see: An Overview of Information Hiding in H.264/AVC Compressed Video