Maier A, Lorch B, Rieß C (2022)
Publication Language: English
Publication Type: Conference contribution, Conference Contribution
Publication year: 2022
URI: https://faui1-files.cs.fau.de/public/publications/mmsec/2022-Maier-BNN-ITNT.pdf
DOI: 10.1109/ITNT55410.2022.9848571
The goal of multimedia forensics is to determine origin and authenticity of images and video. The currently most successful approaches use machine learning, and demonstrate excellent performance in lab settings. However, these methods are still challenged to generalize to images from the internet with potentially complex and partially unknown processing. The current best counter-measure is extensive training data augmentation, but this is extremely costly considering the many possible processing variants that must be covered. In this work, we review and consolidate our recent efforts on a different approach to cope with the challenge of images from unknown provenance. We propose to concentrate the training to the forensic task at hand, and to additionally include a measure for uncertainty to detect when a classifier is not confident on a given input. We believe that uncertainty-aware tools can complement existing efforts when data augmentation fails, and additionally provide valuable feedback to forensic analysts.
APA:
Maier, A., Lorch, B., & Rieß, C. (2022). Bayesian Tools for Reliable Multimedia Forensics. In IEEE (Eds.), Proceedings of the 2022 VIII International Conference on Information Technology and Nanotechnology (ITNT). Samara, RU.
MLA:
Maier, Anatol, Benedikt Lorch, and Christian Rieß. "Bayesian Tools for Reliable Multimedia Forensics." Proceedings of the 2022 VIII International Conference on Information Technology and Nanotechnology (ITNT), Samara Ed. IEEE, 2022.
BibTeX: Download