Source identification in image forensics

Source identification is one of the most important tasks in digital image forensics. In fact, the ability to reliably associate an image with its acquisition device may be crucial both during investigations and before a court of law. For example, one may be interested in proving that a certain photo was taken by his/her camera, in order to claim intellectual property. On the contrary, it may be law enforcement agencies that are interested to trace back the origin of some images, because they violate the law themselves (e.g. do not respect privacy laws), or maybe they point to subjects involved in unlawful and dangerous activities (like terrorism, pedo-pornography, etc). More in general, proving, beyond reasonable doubts, that a photo was taken by a given camera, may be an important element for decisions in court. The key assumption of forensic source identification is that acquisition devices leave traces in the acquired content, and that instances of these traces are specific to the respective (class of) device(s). This kind of traces is present in the so-called device fingerprint. The name stems from the forensic value of human fingerprints. Motivated by the importance of the source identification in digital image forensics community and the need of reliable techniques using device fingerprint, the work developed in the Ph.D. thesis concerns different source identification level, using both feature-based and PRNU-based approach for model and device identification. In addition, it is also shown that counter-forensics methods can easily attack machine learning techniques for image forgery detection. In model identification, an analysis of hand-crafted local features and deep learning ones has been considered for the basic two-class classification problem. In addition, a comparison with the limited knowledge and the blind scenario are presented. Finally, an application of camera model identification on various iris sensor models is conducted. A blind scenario technique that faces the problem of device source identification using the PRNU-based approach is also proposed. With the use of the correlation between single-image sensor noise, a blind two-step source clustering is proposed. In the first step correlation clustering together with ensemble method is used to obtain an initial partition, which is then refined in the second step by means of a Bayesian approach. Experimental results show that this proposal outperforms the state-of-the-art techniques and still give an acceptable performance when considering images downloaded from Facebook