Forensic Modules for the Automated Analysis of Physical Integrity in Images (FENCE)
Abstract
As part of the Defense Advanced Research Project Agency (DARPA) Media Forensics (MediFor) Program this contract was to contribute to the research, development and testing of advanced media forensic techniques to automatically detect and locate media manipulations. A number of techniques were pursued. Three scene based algorithms were researched first. Two techniques for the automatic detection of asymmetric cropping of digital images, one was based on metadata analysis and the second based on exploiting geometrical characteristic of the scene by estimating image projective invariants. The third technique was to detect face splicing in images based on the physical analysis of the imaged scene. While the performance of these scene based methods, under controlled scenarios, has been higher than the state of the art, results in the MediFor Evaluations were not satisfying. The team moved on to developing two novel techniques for unsupervised forensic analysis of video file containers. These two tools achieved the highest Area Under the Curve (AUC) scores in the video manipulations task of MediFor evaluations, while at the same time being among the least computationally expensive algorithms. Lastly, an analysis of Photo Response Non-uniformity (PRNU) on images and videos acquired by smartphones was investigated and in particular countermeasures developed for the challenges smartphones present. These last PRNU techniques were not mature enough to participate in the MediFor evaluations, but are presented in this report. This overall effort contributed positively to the greater MediFor research project and the results are summarized in this report.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 05, 2020
- Accession Number
- AD1114249
Entities
People
- A. Piva
- C. Colombo
- D. Baracchi
- D. Shullani
- M. Luliani
- Marco Fanfani