Generation of Synthetic Latent Fingerprints

The aim of the project is to be able to generate synthetic latent fingerprints to be utilized to improve performances of data-hungry deep-learning based recognition systems.

Fingerprints were observed to be unique to each person centuries ago and since the late 19th century, they have been intensively used for personal security and identification. Although the application methods kept changing according to requirements and emerging technologies through those years, the importance of fingerprint in the functioning of the societies persisted. Human fingertips (in fact also the hand palms and the soles of their feet) are covered with skin ridges that are thought to increase friction for better gripping and amplify vibrations for better transmitting signals to sensory nerves. Fingerprints are imprints that are left on surfaces intentionally under control (reference) or unknowingly (latent) due to those ridges and can later be collected from those surfaces using variable methods.

Latent fingerprints are left on a surface unintentionally and they are normally invisible to the naked eye. However, they can be made clear using physical or chemical methods and collected. This is called lifting. Due to this process, latent prints usually sustain more severe variations and deformations compared to reference prints. Moreover, the surface where the latent is left may have background noise or even another overlapping fingerprint. Latent fingerprints lifted from a crime scene are usually matched to reference fingerprints in the database of the law enforcement units (latent vs. reference comparison). In some cases, latent fingerprints lifted from the same crime scene may be compared with each other to determine multiple prints from the same finger (latent vs. latent comparison).

Because of all these difficulties, latent fingerprint recognition systems have much lower success rates compared to reference fingerprint recognition systems. The biggest obstacle in the way of improving these rates is the scarcity of data. As explained above, collection of latent fingerprints is a demanding task in terms of time and labor. The latents collected from the crime scenes by law enforcement are out of the question since they are legal evidence and sharing them without the consent of the owners is a privacy violation and a security flaw. Therefore, there is no large and extensive latent fingerprint database that is open to public for research purposes. Today, the developments in data-hungry deep-learning based applications make this need much more apparent. Recently, few deep learning based systems to analyze and recognize latents have been proposed. However, reference fingerprint databases are being used to train them.

In this project, it is aimed to generate synthetic latent fingerprints to meet the need elucidated above. Although not many, there are some studies on synthetic fingerprint generation in the literature. But, firstly, these studies focus on reference fingerprints. Additionally, a big portion of these studies intends to put the existing fingerprint recognition systems to extensive tests and another portion aims to generate fingerprints to devise biometric spoofing attacks. The utilization of the synthetically obtained samples to train automatic fingerprint identification systems and measuring their contribution to the performance of these systems are left out of the picture.

In the scope of the project, firstly, existing synthetic fingerprint generation methods will be studied and by using the generated samples for training, their impact on the success rate of a deep learning based fingerprint recognition system will be assessed. Later, if applicable, these methods will be modified to produce latent samples and deep neural network experiments will be repeated with the newly generated data. Finally, in the light of the new information, new synthesizing / simulation methods to produce more realistic and diverse samples will be proposed and experimented with.

A diversified and large latent fingerprint database open for the use of researchers would help accelerating the research in this field and make way for the improvement of automatic identification performances. Due to the presently low success rates, latent fingerprint matching mostly relies on human effort and supervision today. Since both the work force and time is limited, securing the justice may be delayed or even obstructed by the mistakes made. Human factor in the process should be minimized as much as possible, meaning that any small improvement in the performances is crucial.