|Thesis abstract: |
The recent development of multimedia devices and editing tools, together with the proliferation of video sharing web sites, has made the acquisition, alteration, and diffusion of video content relatively easy tasks. As a consequence, we find more and more video sequences available on the Internet, but each of them is potentially tampered with by anyone. It is then clear that the development of tools that enable the recovery of past history of video sequences in order to prove their origin and authenticity is more than an urgent necessity. In the last few years, many forensic techniques have been proposed to detect malicious modifications of multimedia data. However, many solutions are specifically tailored to the still image case. Video solutions have been proposed only more recently. For this reason, in this thesis, we focus on the challenging development of video forensic algorithms capable of detecting specific information about the past history of a video sequence, operating in a complete blind fashion. This means that digital content does not need to be signed beforehand, and the analysis is completely conducted having at our disposal only the final object, with no priors about its past. This is possible due to the fact that any non-invertible operation leaves peculiar traces on the video. Therefore, coding, editing, and other operations are characterized by the footprints they leave behind. The analysis and detection of such footprints allow to reveal if a video has undergone a specific operation. More specifically we can split the proposed algorithms in three categories: i) coding; ii) editing; iii) acquisition. Algorithms belonging to the first category exploit traces left by coding. Using these traces we propose a method to blindly detect how many times a video has been compressed in its lifetime, and which is the first codec used to encode a video that has been double encoded. Algorithms in the editing category exploit traces left by specific operations. First we propose a method to detect if a video (or part of it) has been temporally interpolated. Then we focus on the problem of object insertion and removal, and we show how to detect if such an attack has been operated. Since these detectors can be fooled by anti-forensic techniques, we also study the re-capture problem, which is a simple yet effective anti-forensic operation. Starting from this analysis we develop a detector to reveal if a video has been re-captured. The developed algorithms contribute theoretically and algorithmically to the widening panorama of digital forensics. They are always validated by means of a set of experiments on real video sequences. Indeed, we aim to propose solutions valid in a real world scenario. To this purpose, we have also released some of the generated video sequences to enrich public datasets.