According to a paper presented at a security conference in Los Angeles (PDF), the artificially generated fingerprints, dubbed “DeepMasterPrints” by the researchers from New York University, were able to imitate more than one in five fingerprints in a biometric system that should only have an error rate of one in a thousand.

In order to work, the DeepMasterPrints take advantage of two properties of fingerprint-based authentication systems. The first is that, for ergonomic reasons, most fingerprint readers do not read the entire finger at once, instead imaging whichever part of the finger touches the scanner.

Crucially, such systems do not blend all the partial images in order to compare the full finger against a full record; instead, they simply compare the partial scan against the partial records. That means that an attacker has to match just one of tens or hundreds of saved partial fingerprint in order to be granted access.

The second is that some features of fingerprints are more common than others. That means that a fake print that contains a lot of very common features is more likely to match with other fingerprints than pure chance would suggest.

Based on those insights, the researchers used a common machine learning technique, called a generative adversarial network, to artificially create new fingerprints that matched as many partial fingerprints as possible.

The neural network not only allowed them to create multiple fingerprint images, it also created fakes which look convincingly like a real fingerprint to a human eye – an improvement on a previous technique, which created jagged, right-angled fingerprints that would fool a scanner but not a visual inspection.

They compare the method to a “dictionary attack” against passwords, where a hacker runs a pre-generated list of common passwords against a security system.

Such attacks may not be able to break into any specific account, but when used against accounts at scale, they generate enough successes to be worth the effort.

Alex Hern