In our recent article “Content-Aware Detection of Temporal Metadata Manipulation,” published in IEEE Transactions on Information Forensics and Security (T-IFS), we show how to verify if visual content is consistent with alleged timestamps of images, given an outdoor photograph and its geographical position. We also explore which elements in a scene might indicate timestamp inconsistency. Finally, we present an extension to estimate possible time-of-capture when a timestamp is missing.
Our method considers an end-to-end multi-modal architecture, consisting of a multi-task CNN for satellite and ground imagery and MLPs for time and geolocation information, enriched by learning transient attributes. We achieve SOTA for timestamp verification without strong assumptions on how images were captured.
Can you spot the two inconsistent timestamps in the video below? Our Algorithm can 😊