Tools for the Generation of Synthetic Biometric Sample Data (GENSYNTH)
Final Report Abstract
The project GENSYNTH arose in response to a lack of public datasets of fingerprints that are on the one hand privacy-friendly and on the other hand contain a large number of samples. In view of current regulations for the protection of personal data, such as the General Data Protection Regulation (GDPR) in the EU, it became clear that real fingerprints do not satisfy either of the two aforementioned requirements, because they are neither privacy-friendly nor is it feasible to collect them in very large quantities. In contrast, generating synthetic biometric samples of virtual persons seams to be a solution to this challenge. In this project, we explore which and to what extent modern generative methods are suitable for the synthesis of fingerprint images in the fields of biometrics and forensics. Generative Adversarial Network (GAN) architectures are analyzed and the most suitable ones are selected, extended and used for synthesis. It has been demonstrated that GANs from the StyleGAN family as well as image-to-image translation networks such as pix2pix are able not only to generate realistic fingerprints but also to control identity and person-related characteristics. A bunch of generative models have been trained for both random and identity-aware generation of realistic fingerprints. Next, these models have been applied for compilation of large-scale synthetic fingerprint datasets. The source code, generative models and synthetic datasets can be found on our GitLab website. We also introduce the general requirements on synthetic biometric samples and try to establish a commonly accepted evaluation methodology, which includes the following properties of generated samples: realistic appearance, sufficiently high image resolution, anonymity and diversity, as well as the following properties of generated models: identity-aware generation with the control over personal (privacy-related) attributes, inheritance of visual and statistical characteristics of training samples, and last but not least the fair distribution of personal attributes in generated datasets. Synthetic fingerprints generated by our pix2pix models are shown to be almost indistinguishable from real ones as well as our synthetic datasets are designed to possess characteristics of real biometric datasets so that we believe they can be applied for privacy-friendly testing of fingerprint matching algorithms. For dataset compilation we combine model-based and data-driven synthesis, so that the resulting fingerprints are realistic, anonymous, sufficiently diverse and resemble the samples captured a CrossMatch Verifier 300 fingerpint sensor. A fully controlled environment enables synthesis of not only non-mated but also mated fingerprints.
Publications
-
General Requirements on Synthetic Fingerprint Images for Biometric Authentication and Forensic Investigations. Proceedings of the 2021 ACM Workshop on Information Hiding and Multimedia Security, 93-104. ACM.
Makrushin, Andrey; Kauba, Christof; Kirchgasser, Simon; Seidlitz, Stefan; Kraetzer, Christian; Uhl, Andreas & Dittmann, Jana
-
Generation of Privacy-friendly Datasets of Latent Fingerprint Images using Generative Adversarial Networks. Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 345-352. SCITEPRESS - Science and Technology Publications.
Seidlitz, Stefan; Jürgens, Kris; Makrushin, Andrey; Kraetzer, Christian & Dittmann, Jana
-
On feasibility of GAN-based fingerprint morphing. 2021 IEEE 23rd International Workshop on Multimedia Signal Processing (MMSP), 1-6. IEEE.
Makrushin, Andrey; Trebeljahr, Mark; Seidlitz, Stefan & Dittmann, Jana
-
Data-driven Reconstruction of Fingerprints from Minutiae Maps. 2022 IEEE 24th International Workshop on Multimedia Signal Processing (MMSP), 1-6. IEEE.
Makrushin, Andrey; Mannam, Venkata Srinath; Rao, B.N. Meghana & Dittmann, Jana
-
A Survey on Synthetic Biometrics: Fingerprint, Face, Iris and Vascular Patterns. IEEE Access, 11, 33887-33899.
Makrushin, Andrey; Uhl, Andreas & Dittmann, Jana
-
Data-Driven Fingerprint Reconstruction from Minutiae Based on Real and Synthetic Training Data. Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 229-237. SCITEPRESS - Science and Technology Publications.
Makrushin, Andrey; Mannam, Venkata & Dittmann, Jana
-
Synthetische Daten in der Biometrie. Datenschutz und Datensicherheit - DuD, 47(1), 22-26.
Makrushin, Andrey & Dittmann, Jana
