Feature extraction and machine learning solutions for detecting motion vector data embedding in HEVC videos

Abstract

This paper proposed three feature extraction solutions suitable for detecting data embedding in motion vectors (MVs) of coded HEVC videos. In the first feature extraction solution, videos under consideration are reencoded and features are extracted from both videos from Coding Units (CUs). The difference between the two feature sets form a feature vector at a CU level. The CU level feature vectors are then summarized by computing the average and standard deviation of individual features. This summarization is computed at a frame-level and at a video sequence level. Detection models are then used to detect MV data embedding. To generate the detection models, HEVC videos are reencoded whilst employing two different data embedding solutions. Feature variables are then computed and the detection models at frame and sequence levels are generated. The second solution is an extension of an existing work that uses the concept of MV consistency for computing feature variables. In this work, we extent the MV consistency concept to HEVC coded videos by grouping sub CUs based on their coding depths. One set of features is computed by finding the joint probability that a CU has a given coding depth and the bitrate of the MV differences of the sub CUs are smaller than or equal to the between-CU MV differences. Another set of features is computed by finding a similar probability but for sub CUs with MV differences greater than the between-CU MV differences. The third solution combines the features of all of the aforementioned solutions resulting in a set of 16 feature variables. The feature variables are visualized by projecting them using spectral regression where it is shown that the third solution results in separable features. In comparison to existing work, experimental results show excellent classification accuracies for HEVC videos coded at different spatio-temporal resolutions and different bitrates.