Detecting Manipulated Videos
Videos faked using AI can hardly be distinguished from real ones. But how can a deepfake be recognized and proven as such?
The CyberSec Research and Transfer Center (FTZ) at HAW Hamburg is working together with Chainstep GmbH on a technology that makes it possible to detect manipulated videos. The project "Robust Signature of Audiovisual Media against (Deep) Fakes (SaM-fake)" was successfully presented in the BMBF funding program "DATI Innovation Sprints."
How the Technology Works
The "Trusted Cam" software deposits invisible watermarks throughout the video during recording. These would be at least partially destroyed by a manipulating AI.
"It is almost impossible to completely remove the watermark," explains Prof. Dr. Volker Skwarek, Scientific Director of FTZ CyberSec. "Parts of it will always be recognizable. This can then be used to deduce the manipulation of a video that has already been secured."
Simple image processing such as rotating, cropping, scaling, or even compressing files does not affect the stored watermarks.
Versatile Applications
The technology can be used in many ways – for example, in messenger applications or when sending via social media. In messenger apps, recorded videos and audio can be signed directly with the software. When receiving these signed files, they are automatically verified.
The goal: open source software that can be integrated into a messenger app – for proof of authenticity, for example when reporting from crisis areas.