Research On The Theory And Methods Of Temporal Domain Video Forgery Forensics

With the development of multimedia and internet technology,digital videos become more and more popular in our daily life.However,video editing software makes it easy to tamper with the content of digital videos,leading to a series of information security problems.In practice,video forgery detection techniques are critical to ensure the originality,authenticity and integrity of digital videos,and attract researchers’ attention.Among the various video tampering methods,time domain tampering is most widely employed for simplicity reason.Time domain tampering involves frame deletion,frame duplication and frame insertion.This thesis studied the special characteristics of videos that have been tampered with in time domain and proposed three time domain forgery detection methods.The major contributions of this thesis are summarized as follows:(1)To detect the frame duplication forgery,we proposed a detection algorithm based on the video sub-sequence fingerprints.A video sub-sequence can be seen as a three-dimensional matrix which consists of spatial frames and time axis,with temporal and spatial information at the same time.Frame duplication forgery often causes highly similar sub-sequences,and the similarities will become the key evidences for frame duplication forgery detection.We designed a digital fingerprint for video sub-sequence.We first exploited the temporally informative representative images to fuse the temporal information of a video sub-sequence.Block-based discrete cosine transform is then used to extract the spatial information.Finally the coefficients at special positions are binarized to construct the digital fingerprint carries both spatial and temporal information of a video sub-sequence.We made pair wise comparison of the fingerprints in the whole video to detect whether it had been forged.The exact locations where tampering occurred can also be detected.Experimental results show that the proposed algorithm makes a good tradeoff between detection accuracy and computational complexity.It achieves a good overall detection performance and shows robustness against some common video processing attacks such as luminance change and MPEG compressing.(2)For the surveillance videos taken indoor under fluorescent light that contain lots of static parts,we proposed a time domain video forgery detection algorithm based on electric network frequency(ENF)signal.The lights flash in the videos represents the ENF signal.By extracting the ENF signal and searching for the sudden phase changes of the ENF signal,the locations of forgery would be detected.The key contribution of this work is researching and solving the problems of frequency shifting and not adequate data for ENF signal caused by low frame rates.We designed a band-pass filter according to Nyquist-Shannon sampling theorem to extract the ENF signal from videos.The cubic spline interpolation is employed to reconstruct the ENF signal extracted from videos,which solves the problem of data lacking caused by low sampling rate.The proposed method is one of the earliest methods using the ENF signal in videos independently dispensing with reference signal.Experimental results demonstrate that the proposed algorithm has good performance to detect the time domain forgery in the video captured indoor.For those forged videos that could not be detected by human eyes or other content-based detection methods,our algorithm can still detect the exact location of forgery.(3)For the newest H.265/HEVC compressed videos,we proposed a time domain video forgery detection algorithm based on the features of the coding parameters.With the increasing of the video quality,the demand for storage is rising,and the most advanced video coding standard H.265/HEVC is used widely.In this dissertation,we studied the special features,caused by time domain video forgery,of the compressed domain parameters in H.265/HEVC videos.To detect the exact location of time domain forgery without a complete decoding,we employed features such as the intra prediction pixel ratio in the P frames,the inter prediction pixel ratio and the average length of motion vectors from different directions in the B frames.The proposed method is one of the earliest methods focus on detecting the time domain forgery in H.265/HEVC videos.Experimental results illustrate that the proposed algorithm has good detection performance for the time domain forgery in H.265/HEVC videos.The algorithm also shows strong robustness to the miss hits and false alarms caused by the video content.