Multimedia source identification using an improved weight photo response non-uniformity noise extraction model in short compressed videos

Source detection of digital images and videos is essential to multimedia security forensics research. Due to the defects of the manufacturing process, the optical sensor generates a photo response non-uniformity noise attached to it when generating multimedia content. This noise is widely regarded as the inherent fingerprint of multimedia due to its uniqueness and can be used for multimedia source detection. However, videos shared on social networking sites will be further compressed. It is difficult for existing methods to extract enough PRNU noise from such compressed videos for source identification, especially short compressed videos. Considering the effect of compression on PRNU noise, a weighted PRNU extraction model based on variance-stabilized transformed multi-scale iterative least-squares filtering is proposed. First, the video’s decoder needs to be reconstructed. When the decoder’s loop filter is removed, more PRNU noise is retained in the video frame. Secondly, a multi-scale iterative least-squares filtering algorithm based on the variance stabilized transform (VST) is proposed, which can extract more PRNU noise from the video frames. Finally, maximum likelihood estimation using quantization parameter (QP) weighting enables a better estimate of the PRNU and suppression of other noise. Experiments on the public VISION database show that the proposed PRNU extraction model improves the recognition performance by about 20% on average over existing models.