Neural Network Learns to Forge Fingerprints
Fingerprint identification is considered one of the more reliable ways to verify a person’s identity. Of course, it’s best used in combination with other methods—multi-factor authentication is still the gold standard. Nevertheless, fingerprint technologies are used by software developers and device manufacturers more often than any other biometric method.
However, this method may soon become less secure. In the United States, researchers have developed a neural network capable of forging fingerprints. The computer generates images that are recognized by various sensors as fragments of real human fingerprints.
To train their neural network, the project’s authors used real data from 5,400 people. While this isn’t a huge sample, it was sufficient for training the neural network. The researchers published their results in a preprint article.
Currently, the system can only generate fingerprint fragments—but that’s enough for many systems that use fragments for identification. Even though fingerprints are unique, which is why fingerprint authentication was developed, it’s still possible to fool the system.
Vulnerabilities in Fragment-Based Systems
This is especially relevant for systems that work with fingerprint fragments. It turns out that it’s possible to create an artificial pattern that matches the fingerprints of several people at once. Many fingerprint scanners built into smartphones, laptops, and other electronic devices only use part of a fingerprint. This fragment is stored in a database and compared to the scanned print during authentication.
Some fingerprint fragments have repeating patterns, which allowed the researchers to implement their project.
From MasterPrints to Deep MasterPrints
Previously, a neural network called MasterPrints was developed, which could modify details of existing fingerprints but couldn’t generate entirely new ones. The team from New York University, featured in this article, achieved more. They recently introduced a neural network called Deep MasterPrints, which can generate universal “master keys”—fingerprint fragments based on a given template.
The effectiveness of Deep MasterPrints is impressive: it can generate fragments that match 76% of the sample set. The previous neural network achieved only 33.4%. To avoid errors, the researchers also tested existing random fingerprint generation systems, which matched only about 7% of the control sample—making them ten times less effective than the new neural network.
Levels of Fingerprint Security
Experts distinguish three levels of fingerprint identification security. The highest level has a false match rate of 0.01%, the medium (and most common) level has a 0.1% false match rate, and the lowest level has a 1% false match rate. The New York University team worked with the lowest level. They note that most biometric sensors operate at the 0.1% false match rate, where the system’s effectiveness drops to 23%. The highest security level remains out of reach for the neural network—it was only able to fool ultra-precise sensors in 1.3% of cases.
Regardless, this technology is advancing, and more impressive results can be expected soon. One advantage of Deep MasterPrints is that it works with digital fingerprints, not just a database of raster images of different people’s fingerprints.
According to the researchers, their development will help improve protection methods and user identification systems that rely on fingerprint recognition.
Facial Recognition on the Rise
It’s worth noting that facial recognition technology is increasingly being used instead of fingerprints. For example, Apple’s latest smartphones use FaceID for user identification.