Video steganalysis takes effect when videos corrupted by the target steganography method are available. Nevertheless, classical classifiers deteriorate in the opposite case. This paper presents a method to cope with the problem of steganography method mismatch for the detection of motion vector (MV) based steganography. Firstly, Adding-or-Subtracting-One (AoSO) feature against MV based steganography and Transfer Component Analysis (TCA) for domain adaptation are revisited. Distributions of AoSO feature against various MV based steganography methods are illustrated, followed by the potential effect of TCA based AoSO feature. Finally, experiments are carried out on various cases of steganography method mismatch. Performance results demonstrate that TCA+AoSO feature significantly outperforms AoSO feature, and is more favorable for real-world applications.
Spatio-Temporal Rich Model-Based Video Steganalysis on Cross Sections of Motion Vector Planes