Passive Forensics For The Content Tampering Of Digital Video

Digital media can be easily edited and modified. With the rapid development of various image and video editing software tools, professional or ordinary users can maliciously produce and spread some tampered media for different purposes. As an important measure for information management, video passive forensics technique has become a hot research topic in the field of multimedia information security. Passive video forensics utilizes the inherent features of image and video capturing and/or the left traces during various editing operations to check their authenticity and integrity. It does not need any pre-processing such as digital watermark, and does not need prior information. This makes passive forensics more practical. Wide research attention is paid to passive video forensics. The main works and contributions of this thesis are summarized as follows.First, a video forensics scheme is proposed based on the block-based sensor pattern noise (SPN), whose adaptive-threshold for classification is obtained by maximum likelihood estimation (MLE). It extracts the SPN by wavelet de-noising and Weiner filter. By setting a sliding window of fixed size, block-based energy gradient, signal-noise ratio and the correlation between the SPNs of blocks with the same positions in neighboring frames are computed to build feature vector. MLE is utilized to obtain the adaptive threshold for classification. It is quite different the empirical threshold selection in the literature. Experimental results show that the proposed approach is effective for the forensics of cope-paste based tampering to the contents of digital video. Moreover, it can locate the tampering of small regions in digital video.Second, an adaptive threshold selection scheme is proposed for passive video forensics. It is based on the un-decimated dyadic wavelet transform (DyWT). Based on DyWT, it extracts the low-frequency approximation component and high-frequency detail component to realize the passive forensics. To make the extracted features be translation invariant, a new wavelet transform, i.e, DyWT is used instead of the conventional wavelet transform. For the threshold selection, MLE is still utilized similar to the first work in this thesis. Adaptive thresholds are computed for the high-frequency and low-frequency components, respectively.The test video sequences are selected from the public-open forensics library SULFA provided by the University of Surrey, UK. Experimental results show that the proposed approaches have achieved desirable results.