Classification and clustering of video fingerprints: preliminary results

Many photos and videos are produced and uploaded to the Internet every day. Even though this is a small part of the total, a large amount of them is illegal. Once digital content has been distributed online, it is often difficult to re-associate the photo or video to the device that produced it or to the user who initially shared it. To counter the spread of illegal content, there is a branch of studies called “source camera identification”, which aims to reconnect a photo or video to the device that developed it. The idea behind source camera identification is that each camera, having imperfections that make it unique, gives a digital fingerprint to the content it produces. The noise of a digital content, which represents a variation of intensity that cannot be found in the recorded content, contains the fingerprint along with some random factors. The noises, which are extracted through denoising algorithms, can be used directly to identify the device that produced the content, or they can be used to estimate the fingerprint. This thesis works in the source camera identification of video content. Two datasets are considered: one called Vision, which is considered the reference dataset in this area and one made available by the University of Bologna. The work carried out in this thesis was to extract the noises on those datasets, and calculate the fingerprints, comparing different approaches present in the state of the art. The approach that was chosen has yielded the best results through a classi- fication algorithm. Once the noises were extracted and the fingerprints calculated, classification and clustering techniques were applied. Two classification techniques have been developed one through convolutional neural network and another using a function called Peak-to-correlation energy. Clustering algorithms have been applied, already developed to work in this area, one that considers a known number of cameras and another that considers an unknown number.