Research On Passive Forensics Of Video Motion-compensation Frame-interpolation

With the availability of inexpensive and portable video capture devices,digital video is enriching our daily life.Meanwhile,the proliferation of powerful video editing tools makes it much easier than ever to produce faked videos without leaving any perceptible traces.This breaks our traditional concept of “seeing is believing”and brings serious crises with respect to public confidence.Video forensics,which attempts to verify the authenticity and integrity of digital videos,has attracted wide research interests in the field of information security.Especially,passive forensics approaches,which detect tampering traces without the aid of any prior axillary information,are extensively studied in recent years.Among various video editing operations,motion-compensated frame-interpolation(MCFI)is a special frame based video manipulation.It is widely used in video and film production,which periodically synthesizes interpolation frames from adjacent reference frames to increase the frame-rate.Though MCFI is originally proposed to improve the visual quality of low frame rate videos,it can also be utilized by counterfeiters to fake high frame-rate video or splicing videos with different frame-rates for malicious video forgeries.This dissertation belongs to the field of passive video forensics,and focuses on the artifacts/traces caused by MCFI for its forensics analysis.The main works and contributions are summarized as follows:(1)A residual model is theoretically built to express the MCFI based video frame-rate tampering operation.After analyzing the principles behind various MCFI techniques,the residual is stemmed from the difference of interpolated frame and its absent original frame.Subsequently,it is verified from the quantitative and qualitative perspectives.Therefore,this residual model provides us the theoretical support to design interpolation frame localization algorithms.(2)Residuals caused by various MCFI techniques follow Laplace distributions with different variances due to different ME and distinct MCI strategy.Therefore,residual is firstly exploited as forgery trace to expose MCFI,and thus the identification of various MCFI techniques is converted into a problem of discriminating the differences of residual signals inside interpolated frames.Spatial and temporal Markov statistical features are designed to capture the differences of residual signals to identify the adopted MCFI techniques.A pre-classifier,which includes scene change detection,static scene detection,and multi-loop detection method,is designed to improve detection accuracy by suppressing the side effects of static interpolated frames and original frames.Experimental results demonstrate that no matter interpolated frames come from uncompressed videos or compressed videos with high perceptual quality,the proposed method can identify the adopted MCFI technique.(3)An irregularity of optical flow based MCFI forensics algorithm is proposed.Exiting MCFI detectors are design from the change of pixel value of video frame perspectives,and can not consider the subtle change of motion vectors in motion and texture-rich regions of interpolated frame.A novel structural feature is extracted as the three-consecutive frame based temporal difference weighted histogram of local binary pattern(LBP)calculated on the optical flow,which is effective to describe the local change of pixel domain and optical flow domain.Experimental results verify that the proposed algorithm can achieve higher localization results compared with exiting MCFI detectors.(4)A detection approach is proposed for detecting the possibility of unknown MCFI technique.Since it is inferred from residual model that any MCFI technique can cause the change of pixel values in motion and texture-rich regions of interpolated frame,inherent statistics of motion region of interpolated frames by various MCFI techniques inevitably exist difference from that of original frames.To reduce interference from non-motion and motion regions' s information for a test frame,the statistical moments from the motion regions of motion-aligning frame difference are extract.Subsequently,these features are fed into the one-class SVM to perform MCFI detection.Experimental results validate that proposed approach can identify MCFI forgery in a more practical scenarios where the MCFI methods and video sources to be tested are unknown for the forensic analyst.(5)An artifact indicated map(AIM)and Tchebichef moments(TMs)based robust interpolation-frames localization algorithm is proposed.By theoretical analysis and experimental verification,the artifact regions in interpolated frames are closely correlated with high residual energies.Thus,an AIM is designed to select target regions with high residual energies.Meanwhile,blurring effects or deformed structures of artifact regions cause shape change of interpolated frames,the degree of which are quite different from that of theirs neighboring reference frames.This kind of shape changes can be captured by high-order TMs.Specifically,the mean values of absolute high-order TMs are extracted from such regions for forensics analysis.A sliding window scheme is proposed to identify interpolated frames by using the extracted TMs along the temporal direction.The periodicity of frame interpolation is further checked to exclude abnormal frames,which have been wrongly detected due to blurring or unexpected noise.Experimental results manifest that the proposed approach is robust to lossy compression,blur and noises.