Pierluigi Paganini
August 18, 2024
A team of researchers from US universities demonstrated how to deceive fingerprint recognition systems through dictionary attacks using ‘MasterPrints,’ which are fingerprints that can match multiple other prints.
The experts introduced DeepMasterPrints, which are complete image-level prints, and used a method called Latent Variable Evolution, involving a Generative Adversarial Network (GAN) trained on real fingerprint images and a search strategy to generate prints that maximize impostor matches.
Fingerprint recognition is increasingly used in various applications, but small-sized sensors, like those on smartphones, only capture partial fingerprints, making them more susceptible to incorrect matches. The researchers pointed out that their study is the first to create image-level synthetic “MasterPrints,” which can exploit this vulnerability. The research demonstrates that these “DeepMasterPrints” can spoof 23% of subjects at a 0.1% false match rate, and up to 77% at a 1% false match rate. This highlights the significant security risks posed by using small, low-resolution fingerprint sensors.
The researchers trained two generator networks using the Wasserstein GAN (WGAN) algorithm to create synthetic fingerprints. One network was trained on fingerprints from a capacitive sensor, and the other on inked and rolled fingerprints. Both networks used a deep convolutional GAN architecture and were trained adversarially with a Wasserstein loss function and RMSProp optimizer at a learning rate of 0.00005. Each generator was trained for 120,000 updates, with the discriminator trained five times for each generator update. To prevent blocky artifacts, the researchers switched from deconvolutions to using upsampling with convolutions.
To create a DeepMasterPrint, the researchers needed to optimize the latent variables (inputs) for the generator network. These variables exist in a complex, 100-dimensional space, and the goal was to find the best values (or “points”) in this space that would produce the most effective fingerprint images. The Latent Variable Evolution (LVE) technique samples these points, converts them into images, and scores them based on how many identities they can match. The researchers used an evolutionary algorithm, e.g. CMA-ES, to navigate the challenging optimization task because the process lacks clear gradients. They spent three days per fingerprint in the optimization process, evaluating each one against multiple fingerprint matchers, including the widely used VeriFinger, Bozorth3, and Innovatrics systems. The key weakness exploited by DeepMasterPrints is that a match is considered valid if even one of 12 partial fingerprints matches, making the system vulnerable.
The experts tested DeepMasterPrint attacks mainly against smartphones because of their small fingerprint sensors. Since smartphones use capacitive sensors, the researchers created their DeepMasterPrints using a capacitive fingerprint dataset and evaluated them with the VeriFinger matcher.
“In our work, we created a DeepMasterPrint that is intended to spoof an arbitrary identity in a single try. Previous work had much worse results when given only a single attempt. Besides providing an image, LVE creates a much more effective MasterPrint.” concludes the researchers. “Table 3 has the results of the minutiae-only approaches and the capacitive DeepMasterPrint image [23]. In the previous work by Roy et al. [25], the authors generated a suite of five fingerprint templates that were used sequentially to launch an attack, assuming five attempts. Our results for a single DeepMasterPrint is comparable to this suite of multiple MasterPrints. We expect LVE to do very well in creating sequential DeepMasterPrints.”
The study conducted by the researchers demonstrated that using LVE to find latent variables is possible to produce images matching a large number of fingerprints. The method successfully creates full fingerprint images, which could be used in real attacks. The technique is effective across different fingerprint matchers and datasets and has potential applications in both security and computational creativity research.
Follow me on Twitter: @securityaffairs and Facebook and Mastodon
(SecurityAffairs – hacking, MasterPrints)
