Biometrie-Evaluations-Zentrum (BEZ)

Innovative approaches in biometrics: Generating synthetic data


The research project “Synthetic Characters” is dedicated to the development and evaluation of biometric systems for which large, diverse and realistic data sets are essential. Traditionally, such systems are based on real biometric data, but their availability is often limited due to cost, time and data protection aspects. In addition, they often exhibit bias in terms of gender or ethnicity and usually do not contain detailed ground truth information on characteristics such as pose or lighting conditions.
The “Synthetic Characters” project takes an innovative approach: instead of relying on real data, synthetically generated data sets are used. These allow the flexible and controlled generation of biometric characteristics and therefore offer promising potential for research. The key questions here are: How can large synthetic databases be created efficiently? Can synthetic data replace real biometric data on an equal footing? And how can the quality of synthetic data be reliably measured?

Two main approaches are being investigated here: On the one hand, neural networks, such as StyleGAN2, which enable fast and realistic data generation, but offer little scope for the targeted adaptation of parameters. On the other hand, character generators from the field of computer games, such as the “Meta Human Creator” or the “Character Creator 4.0”. These tools offer a high degree of parameterizability and enable the targeted manipulation of individual parameters, for example for bias analyses. However, the generation process is time-consuming and has not yet reached the level of detail of neural networks.
The project is still in its infancy, but has great potential to take biometric systems to a new level by using synthetic data. It offers a solution to the scarcity of real data and at the same time opens up new ways to improve the fairness and accuracy of biometric technologies.