Camera identification for very low bit rate time varying quantization noise videos

This paper addresses the camera identification based on very low bit rate videos with time varying overall noise pattern statistics. First, the overall noise pattern of each frame of the videos is resized to a column vector. It is found that the elements of the resized vectors approximately follow the Laplace distribution. Hence, the second, the fourth and the sixth order statistic moments of each resized vector are computed and these statistic moments form a new vector. The statistic moment vectors are different at different frames because the overall noise pattern statistics are time varying. Second, the principal component analysis is performed for reducing the total number of features for the camera identification. In particular, the statistic moment vector of each frame is projected to the most major component. The projected components of all the frames form a feature vector for each video. Third, the linear discriminant analysis is performed to minimize the intraclass separation and maximize the interclass separation. However, as many eigenvalues of the interclass separation matrix are close to zero, this optimization problem is severely ill-posed. To address this difficulty, this paper proposes to find the columns span the null spaces of the corresponding matrices. The feature vector of each video is projected to these columns and forms a new vector. It is found that these new vectors are pairwisely linear separable. Hence, a set of perceptrons can be employed for the camera identification. Computer numerical simulations show that our proposed method significantly outperforms the conventional method based on computing the correlation coefficients on the photo response nonuniformity noise (PRNU) and the conventional 1-nearest neighbor rule approach in terms of both the identification rate and the robustness performance.