We face an abundance of digital imagery, as well as the wide availability of photo editing software. Knowing that a digital image is authentic becomes increasingly important in many scenarios, for instance law enforcement, journalism and for insurance companies. In past times, human experts were called to assess images from dubious sources. Digital imagery opens novel opportunities, as well as new challenges in determining authenticity and origin. For instance, statistical artifacts, as well as physical properties like the lighting environment are very difficult to assess for humans. Computational methods can greatly support the expert in judging images under examination. Moreover, a number of methodologies can operate in a fully unsupervised manner, such that even large archives can be scanned for manipulated images.

A blind image forensics toolbox should consist of a number of diverse algorithms to detect tampered images. So far, manipulations can often be disguised against a particular detection algorithm, but with an increasing number of tools, it becomes less and less likely that a falsifier manages to evade all detection methods.

My main interest lies in the application of computer vision methodologies to image forensics. I am working on estimating properties of the illumination environment, and verifying the expected physical behavior of natural images.

Check out our project page for a full list of our work, as well as my publications on image forensics!

In a three-year project, we examined ways to optimize real-time banknote sorting using insights from sensor data fusion. Together with our industry partner Giesecke+Devrient, fusion principles have been examined for sensors based on diverse physical measurement principles. The project page is here.