DeepMasterPrints: ‘master key’ fingerprints made by a neural network can now fake fingerprints

2 min read

New York University researchers have found a way to generate artificial fingerprints that can be used to create fake fingerprints. They do this by using a neural network. They have presented their work in a paper titled DeepMasterPrints: Generating MasterPrints for Dictionary Attacks via Latent Variable Evolution.

The vulnerability in fingerprint sensors

Fingerprint recognition systems are vulnerable to dictionary attacks based on MasterPrint. MasterPrints are like master keys that can match with a large number of fingerprints. Such work was done previously at feature level, but now this work dubbed as DeepMasterPrints has much higher attack accuracy with the capacity to generate complete images.

The method demonstrated in the paper is Latent Variable Evolution which is based on training a Generative Adversarial Network (GAN) on a set of real fingerprint images. Then a stochastic search is then used to search for latent input variables to the generator network. This can increase the accuracy of impostor matches assessed by a fingerprint recognizer.

Small fingerprint sensors pose a risk

Aditi Roy, one of the authors of the paper exploited an observation. Smartphones have small areas for fingerprint recording and recognition. Hence the whole fingerprint is not recorded in them at once, they are partially recorded and authenticated. Also, some features among fingerprints are more common than others. She then demonstrated that MasterPrints can be obtained from real fingerprint images or be synthesized.

With this exploit, 23% of the subjects could be spoofed in the used dataset at a 0.1% false match rate. The generated DeepMasterPrints was able to spoof 77% of the subjects at a 1% false match rate. This shows the danger of using small fingerprint sensors.

For a DeepMasterPrint a synthetic fingerprint image needed to be created that can fool a fingerprint matcher. A condition was that the matcher should also match that fingerprint image to different identities in addition to realizing that the image is a fingerprint. The paper presents a method for creating DeepMasterPrint using a neural network that learns to generate fingerprint images.

A Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is used for searching the input space of the trained neural network. The ideal fingerprint image is then selected.

Conclusion

Partial fingerprint images can be generated that can be used for launching dictionary attacks against a fingerprint verification system. A GAN network is trained over a dataset of fingerprints, then LVE searches the latent variables of the generator network for a fingerprint image that maximize the matching chance. This matching is only successful when a large number of different identities are involved, meaning specific individual attacks are not so likely. The use of inked images and sensor images show that the system is robust and independent of artifacts and datasets.

For more details, read the research paper.