Duration: April 2021 - March 2024
The goal of the project is to identify and classify attacks and forgeries of images and video data streams using different types of fake identities (deep fakes) in image- and video-based authentication processes and to develop algorithms for their detection. In order to be able to recognize and soundly evaluate attack scenarios and fakes, the project generates new insights into the diverse motivations and backgrounds (from fraud to manipulation of political decision-making processes) for faking images and video data streams, their detection and explanation, and how police authorities and the judiciary deal with such attacks and fakes.
Fraunhofer HHI will mainly contribute to the technical implementation and development of new methods for the generation and detection of deep fakes. The core objective here will be new detectors that robustly detect deep fakes via AI-based methods, exploiting a wide range of information such as possible artifacts, inconsistencies in lighting, facial expressions, movement or vital parameters such as pulse.
In this context, the development of interpretable methods with explainability of the neural network decisions is expected to increase the usability of the results among end users. In order to be able to train and evaluate the new detectors, HHI will also build up a reference database of deep fakes. For this purpose, existing databases and methods will be analyzed and used, but also own improved methods for synthesis and manipulation will be developed in order to improve the robustness of the detectors against future attacks.
- Bundesdruckerei GmbH
- Otto-von-Guericke-Universität Magdeburg
- Fraunhofer Heinrich Hertz Institut
- Hochschule für Wirtschaft und Recht Berlin
- BioID GmbH