Detecting Re-captured Videos using Shot-Based Photo Response Non-Uniformity Dae-Jin Jung
Recent digital camcorders Advantages High quality Low price Easy usage Abuse Camcorder theft Introduction 2
Camcorder theft (illegally re-captured videos) Single largest source of [1] Fake DVDs Unauthorized copies Causes a great loss on movie industry Introduction OriginalRecaptured [1] Motion Picture Association Of America ( 3
Lee et al. [2] Watermarking scheme Robust against camcorder theft Estimates the position of the pirate Good results Needs embedding process Previous Works 4 [2] Digital cinema watermarking for estimating the position of the pirate (2010)
Cao et al. [3] Identifies recaptured images on LCD screens Good results (EER lower than 0.5%) Used SVM Not suitable for videos Previous Works 5 [3] Identification of recaptured photographs on LCD screens (2010)
Wang et al. [4] Detects re-projected video Skew estimating Can achieve low false positive Using many feature points Feature points not on the right position Manual pre-processing is needed Previous Works 6 [4] Detecting Re-Projected Video (2008)
Recording device Original Analog cameras Mainly used in movie industry High quality, soft shades of colors Recaptured Digital cameras Small, light, easy to handle Recapturing without being observed Differences (Original/Recaptured) 7
Number of cameras used in recording Original Many cameras Conversation scenes Different purposes Shots have different source cameras Recaptured Only 1 camera for recapturing Differences (Original/Recaptured) 8
Different post-processing Original Heavy post-processing Harmonize shots from different cameras CGs, visual effects Recaptured Minimum post-processing Resizing Re-compression Differences (Original/Recaptured) 9
Shot based PRNU estimated from an original video Has low correlation with each other Analog camera Many cameras in recording Heavy post-processing Shot based PRNU estimated from a recaptured video Has high correlation with each other Digital camcorder (PRNU) 1 recording camera Light post-processing Resulting characteristics 10
Overview Divide a video into shots Estimate PRNU PCE based recaptured video detection Proposed method 11
Proposed method 12 [5] Automatic partitioning of full-motion video (1993)
PRNU estimation [6] PRNU model MLE method Codec noise removal Proposed method 13 [6] Source digital camcorder identification using sensor photo response non-uniformity (2008)
Proposed method 14
Detecting re-captured videos False negative correction No fine reference pattern from sky view Warshall’s algorithm Proposed method
Test set 10 original videos 20 shots were extracted Full HD ~ HD 4 Digital camcorders Samsung : 1 (H205BD) Sony : 3 (CX500, CX550, SR10) 40 recaptured videos Experimental results 16
Test set Experimental results 17
Re-captured video detection test ( number of true values/total ) ratio in boolean matrix ‘1.00’ indicates a recaptured video Experimental results 18 Recaptured videos
Compression test Quality factor(QF) : 100~60 MPEG4 (AVC/H.264) Experimental results 19
Resize test Scaling factor (SF) : 0.9~0.3 MPEG4 (AVC/H.264) Experimental results 20
Combinational test Common setting for re-compression Quality factor (QF) : 80 Scaling factor (SF) : 0.5 MPEG4 (AVC/H.264) 100% detected Experimental results 21
Automatic recaptured video detection Uses the shot based PRNU Good results Recompressed Resized Still weak against severe attacks Conclusion 22
Thank you
Threshold setting 2400 pairs of PRNU from same camcorders 2400 pairs of PRNU from different camcorders Threshold : 80 Appendix
Un-correctable False negative Appendix