Fingerprint Synthesis Using Deep Generative Models
THESIS STUDENTS:
Diego Figueroa
From: Lima, Peru
Studies: Master’s degree in Machine Learning, Systems and Control at Lund University
Weizhong Tang
From: China
Studies: Master’s degree in Machine Learning, Systems and Control at Lund University
How did you become aware of Precise Biometrics, and what attracted you to do your project with us?
Diego: We met Precise Biometrics at an AI career event held by Lund University about master’s thesis proposals. My thesis partner and I were looking for proposals that could be done in pairs and had a focus on image analysis with machine learning methods. Precise’s thesis proposal had just what we wanted.
What are the main research questions you aim to answer through your thesis?
There are three main research questions we’re aiming to answer:
- Which deep generative models* perform the best in normal fingerprints synthesis?
- Which deep generative models can achieve fingerprint-to-fingerprint transformation in high quality?
- How to measure the quality of generated fingerprints in practice and in statistics?
How do you believe your thesis contributes to the existing knowledge in AI and biometric technology?
Weizhong: It might be a bit bold to talk about how much contribution our thesis can make to AI and biometric fields. I would rather say that our thesis will be able to benefit the company in terms of having a comprehensive insight into the feasibility of fingerprint synthesis*, and ideally generating unique fingerprints in high quality by constructing various ML models*. However, there indeed is a creative point that we have implemented state-of-the-art diffusion models to generate fingerprints, and this has not been done by other researchers in AI or biometric field.
In what ways do you believe your research could be expanded or improved upon in the future?
Diego: I believe that as time passes, more state-of-the-art models will appear, and it will be possible to obtain the same results with less data and parameters. An expanded project would be a synthetic fingerprint pipeline that can be used to create a variety of conditions (normal, dry, spoof).
Weizhong: We have tried a few model structures to generate normal fingerprints or do a fingerprint style transformation. And obviously there are still a bunch of models that we haven’t had time to go through. Other than that, the evaluation of synthetic fingerprints is a critical topic here as well. We may think of how to create or construct a metric so as to assess the quality of the images accurately and statistically.
What are the practical implications of your research findings, and how do you envision them being applied in real-world settings?
Diego: It would be possible to expand the current fingerprint dataset of Precise. This would allow for better training of matching and spoof detection* algorithms.
Weizhong: We have offered a few ways to generate synthetic fingerprints as well as evaluate their quality, which can then be possibly used in real-world for enriching the company’s dataset for each sensor, and ultimately improving the accuracy of fingerprint recognition.
GLOSSARY*
Fingerprint synthesis is the process of generating artificial fingerprint images that resemble real fingerprints and can be used to augment fingerprint datasets for training and evaluating biometric systems.
Machine learning (ML) models are computer algorithms that can learn patterns and relationships in data, and use this knowledge to make predictions or decisions without being explicitly programmed.
Deep generative models are a type of machine learning models that can generate new data samples similar to the training data. These models (e.g., Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models (DDPMs)) are capable of capturing the underlying distribution of the training data and generating new samples that share similar characteristics.
Generative Adversarial Networks (GAN) Models are a type of machine learning model that uses two neural networks, one to generate new data that resembles the training data and another to discriminate between the generated and real data, which work together to improve the quality of the generated data over time.
Spoof detection is the process of identifying whether biometric data, such as fingerprints or faces, is authentic or fake.
Want to write your Master’s Thesis with us?