Source identification of captured video using photo response non-uniformity noise pattern and svm classifiers

Abstract

Recent works have shown that passive capturing source detection methods based on Photo-Response-Non-Uniformity (PRNU) extraction are the most reliable ones in comparison with techniques that based on lens properties or compression artifacts. Some important issues in this field include: employing an effective method for extracting PRNU, calculating the similarity and categorizing videos according to source of camera. In this study, a comprehensive algorithm is proposed to compare and evaluate the performance of different source detection methods in terms of filters used and partitioning process applied for PRNU extraction coupled with SVM classifier. Moreover, in consideration of observations, a new method is proposed for sampling selection using SVM classifier. Furthermore, the capabilities of employing and combining the results of different color parts of videos are used instead of changing them to grayscale. The proposed algorithm is based on three essential steps: Firstly, fingerprint of each camera, which is regarded as reference PRNU, is calculated by extracting PRNU of blue-sky videos. Secondly, the PRNU similarities of sample videos with reference PRNU are measured by calculating cross correlation and Peak to Correlation Energy (PCE) metrics. Finally, the sample videos are classified based on calculated PCE with SVM classifier. Experimental results revealed that Zero-mean and Wiener filters have small influences on PRNU, thus they can be ignored. Experimental results also revealed that eliminating the partitioning step considerably increases the performance of detection success rate by 15%. Among SVM classifiers, “RBF” and “MLP” types have the best identification rate of 75%.