We can gather data re- lated to a specific place, ob ject or event that were shared or posted in vari- ous places (e.g., social net- works, CCTVs and body cams) or that were seized in an operation.
The main task consists of tying together different information pieces in a spatial and temporal coherent form (X-coherence). Space-time-coherent pieces can be exploited to have an overall idea of the event, place or object as a whole, as well as of specific points within the event’s timeline.
With X-coherent pieces, we can (i) look for specific clues (authorship, related topics); (ii) mine possible suspects, objects or places of interest through biometrics and machine learning techniques; (iii) counter fake propaganda and media repurposing through provenance analysis methods; and (iv) assess sensitive content through diverse filters (e.g., violence), everything underpinned by open-set recognition techniques.
This video presents the main ideas behind the DéjàVu project.
Feature-space-time coherence, i.e., space-time coherence in physical terms (position in time), that is where and when something happened.
Given any of these events, it would be extremely important to find associated information pieces in space and time both in the pool of apprehended materials and in possible collected social media information.
Leveraging the feature- space-time coherent information pieces from previous stages to find possible persons, objects or places involved with the event and, ultimately, propose some candidate suspects for further inves- tigation.
Forensics analysts are interested not only in determining if a digital object is fake or real but also in pinpointing who created it, what happened, when and how (genealogy) an asset was created.
Given data collected and processed with the previous steps of the research, we will focus on deciding whether or not an image or video stream presents sensitive content.
Graph models for machine learning: K-associated graphs and Attribute-based Decision Graphs by Prof. João Bertini (left) and Reverse engineering of video content for forensic analysis: model-based and data-driven approaches by Prof. Paolo Bestagini (right) … Continue reading
In this article, the authors present a fully automated large-scale end-to-end pipeline that starts with the step of provenance image filtering (over millions of images) and ends up with the provenance graphs. The most immediate application … Continue reading
DéjàVu was featured on Record TV dominical program Domingo Espetacular. Record TV is among the most popular channel in the Brazilian free-to-air television network. The video (in Portuguese) can be watched here.
