Synthetic Data Generation and Detection

Current advances in artificial intelligence make it possible to create deceptively realistic image, video and audio material. The possible applications are diverse and range, for example, from completely computer-generated photos, to image processing to make night shots look like pictures taken during the day, to imitating the voice of other people.

In so-called deepfakes, the face of one person is projected onto another to create videos of scenes that never took place. Although these tools are often used for entertainment purposes, they also offer considerable potential for abuse. In the age of social media, false information can spread rapidly and uncontrollably, so the dangers of deceptively real deepfakes range from political manipulation, blackmail and defamation to various scams.

In order to prevent these dangers, the research project deals with the detection of artificially created or manipulated audio, image and video material. One of the most common approaches for the development of robust detection algorithms is the combination of two components. According to the adversarial approach, two components compete against each other: A generator creates artificial data, while a classifier tries to distinguish it from real data. In the course of this training process, the results of both the generator and the classifier steadily increase in quality. The aim is to generate synthetic GDPR-compliant test data sets in order to improve existing detection methods with their help.

The primary focus of the project is on the development of a demonstrator for the detection of manipulated images and videos. In the course of this, existing state-of-the-art approaches will not only be applied, but also optimized.