Tracing images back to their social network of origin: A CNN-based approach

Recovering information about the history of a digital content, such as an image or a video, can be strategic to address an investigation from the early stages. Storage devices, smartphones and PCs, belonging to a suspect, are usually confiscated as soon as a warrant is issued. Any multimedia content found is analyzed in depth, in order to trace back its provenance and, if possible, its original source. This is particularly important when dealing with social networks, where most of the user-generated photos and videos are uploaded and shared daily. Being able to discern if images are downloaded from a social network or directly captured by a digital camera, can be crucial in leading consecutive investigations. In this paper, we propose a novel method based on convolutional neural networks (CNN) to determine the image provenance, whether it originates from a social network, a messaging application or directly from a photocamera. By considering only the visual content, the method works irrespective of an eventual manipulation of metadata performed by an attacker. We have tested the proposed technique on three publicly available datasets of images downloaded from seven popular social networks, obtaining state-of-the-art results.