Tools for the Generation of Synthetic Biometric Sample Data (GENSYNTH)

目前，生物识别和数字化取证研究面临着一个严重阻碍这些安全相关领域进展的问题：需要大规模的生物识别数据集来进行灵活和及时的评估，但由于各种原因，其中包括隐私问题，这些数据集都缺失了。后者随着欧盟GDPR的实施而增加，以至于即使是像美国NIST这样成熟的标准化机构也在2018年5月GDPR生效之前删除了大部分公开可用的数据集。为了解决这个问题并解决附带的数据质量维度（定量和定性问题），我们将研究允许生成大规模可信和现实的合成数据集的方法，以实现可重复、灵活和及时的生物识别和法医实验评估，不仅符合我们用现代技术看到的对数据的渴望，而且符合欧盟数据保护立法。为了实现我们的目标，本项目的工作遵循两种不同的解决方案：第一种（数据适应）采用现有的生物识别/法医样本，对它们进行调整以反映某些获取条件（感官、生理和环境可变性），并且（如果应用程序上下文需要）对隐私属性进行上下文敏感控制。第二种方法（合成）根据特定的感官、生理和环境可变性，从零开始制造完全人工的样品。该项目的实际工作重点是数字化法医（潜在）指纹，以及两种生物识别模式指纹（FP）和手和手指的血管数据（即手和手指静脉图像）（HFV）。理论和方法概念以及实证研究结果将被概括，以讨论对其他模式（特别是面部识别）进行研究的潜在好处。

Current day biometric recognition and digitized forensics research struggles with a problem severely impeding progress in these security relevant fields: Large scale datasets of biometric data would be required to allow for flexible and timely assessments, but these are missing due to various reasons, amongst them privacy concerns. The latter have increased with the EU GDPR to an extend that even well established standardization bodies like NIST in the USA removed a large part of their publically available datasets before the GDPR became effective in May 2018.To solve this problem and address the attached data quality dimensions (quantitative as well as qualitative concerns), we will research methods allowing for the generation of large-scale sets of plausible and realistic synthetic data to enable reproducible, flexible and timely biometric and forensic experimental assessments, not only compliant with the hunger for data we see with modern day techniques, but also with EU data protection legislation. To achieve our goals, the work in this project follows two distinct solution approaches: The first (data adaptation) takes existing biometric / forensic samples, adapts them to reflect certain acquisition conditions (sensorial, physiological as well as environmental variability), and (if required by the application context) conducts context sensitive control of privacy attributes. The second approach (synthesizing) creates completely artificial samples from scratch according to specified sensorial, physiological as well as environmental variability.The practical work in the project is focused on digitized forensic (latent) fingerprints as well as on the two biometric modalities fingerprint (FP) and vascular data of hand and fingers (i.e. hand- and finger-vein images) (HFV). The theoretical and methodological concepts and empirical findings will be generalized, to discuss the potential benefits of the research performed also for other modalities (esp. in face recognition).