Temporal image forensic analysis for picture dating with deep learning

Estimating the acquisition date of digital photographs is crucial in image forensics. The task of dating images by processing their contents should be reasonably accurate as this can be used in court to resolve high profile cases. The goal of temporal forensics analysis is to find out the links in time between two or more pieces of evidence. In this paper, the problem of picture dating is addressed from a machine learning perspective, precisely, by adopting a deep learning approach for the first time in temporal image forensics. In this work, the acquisition time of digital images is estimated in such a way that the analyst can identify the timeline of unknown digital photographs given a set of pictures from the same source whose time ordering is known. By applying Convolutional Neural Networks (CNN), namely the AlexNet and GoogLeNet architectures in both feature extraction and transfer learning modes, results have shown that the networks can successfully learn the temporal changes in the content of the digital pictures that are acquired from the same source. Interestingly, although images are divided into non-overlapping blocks in order to increase the number of training samples and feed CNNs, the obtained estimation accuracy has been from 80% to 88%. This suggests that the temporal changes in image contents, modelled by CNNs, are not dependent on block location. This has been demonstrated on a new database called 'Northumbria Temporal Image Forensics' (NTIF) database which has been made publicly available for researchers in image forensics. NTIF is the first public database that captures a large number of images at different timeslots on a regular basis using 10 different digital cameras. This will serve the research community as a solid ground for research on picture dating and other image forensics applications.