“Do you think that every fingerprint is unique?” It’s a question a professor posed to Gabe Guo during a casual conversation while he was stuck at home during the COVID-19 lockdowns, waiting to begin his freshman year at Columbia. “Little did I know that conversation would set the stage for the focus of my life for the next three years,” Guo said
Guo, an undergraduate senior in Columbia’s computer science department, led a team that conducted research on the subject, with professor Wenyao Xu of the University of Buffalo serving as one of his co-authors. The paper, published this week in the journal Science Advances, appears to call into question a long-held belief about fingerprints: they are not all unique, according to Guo and his colleagues.
AI-powered research questions the uniqueness of fingerprints
Journals rejected the work several times before the team appealed and had it accepted by Science Advances. “There was a lot of pushback from the forensics community at first,” recalled Guo, who had no forensics experience before the study.
“For the first iteration or two of our paper, they said it’s a well-known fact that no two fingerprints are alike. I guess that helped to improve our study because we just kept putting more data into it (increasing accuracy) until eventually the evidence was incontrovertible,” he said.
As it worked, the AI-based system discovered that fingerprints from different fingers of the same person shared strong similarities and was thus able to tell when the fingerprints belonged to the same individual and when they didn’t, with a single pair’s accuracy peaking at 77%, seemingly disproving that each fingerprint is “unique.”
“We found a rigorous explanation for why this is the case: the angles and curvatures at the center of the fingerprint,” Guo said.
For hundreds of years of forensic analysis, he explained, people have looked at different features known as “minutiae,” which are the branchings and endpoints in fingerprint ridges that are used as traditional markers for fingerprint identification. “They are great for fingerprint matching but not reliable for finding correlations among fingerprints from the same person,” Guo told me. “And that’s the insight we had.”
The authors are confident in their findings and have open-sourced the AI code for others to review
The authors acknowledged that the data may contain biases. According to the study, while they believe the AI system operates similarly across genders and races, more careful validation is required through the analysis of a larger and broader database of fingerprints before the system can be used in actual forensics.
However, Guo believes the discovery will help criminal investigations.
“The most immediate application is that it can help generate new leads for cold cases where the fingerprints left at the crime scene are from different fingers than those on file,” he said. “But on the flip side, this won’t just help catch more criminals. This will also actually help innocent people who might not have to be unnecessarily investigated anymore. And I think that’s a win for society.”
According to Christophe Champod, a forensic science professor at the University of Lausanne’s School of Criminal Justice in Switzerland, applying deep learning techniques to fingerprint images is an intriguing topic. However, Champod, who was not involved in the study, stated that he does not believe the research revealed anything new.
“Their argument that these shapes are somewhat correlated between fingers has been known from the early start of fingerprinting when it was done manually, and it has been documented for years,” he said. “I think they have oversold their paper, by lack of knowledge, in my view. I’m happy that they have rediscovered something known, but essentially, it’s a tempest in a teacup.”
In response, Guo stated that no one had ever systematically quantified or applied the similarities between fingerprints from different fingers of the same person to the extent that the new study has.
“We are the first to explicitly point out that the similarity is due to the ridge orientation at the center of the fingerprint,” Guo said. “Furthermore, we are the first to attempt to match fingerprints from different fingers of the same person, at least with an automated system.”
Simon Cole, a professor in the Department of Criminology, Law, and Society at the University of California, Irvine, agreed that the paper is intriguing but that its practical utility is exaggerated. Cole was also not a participant in the study.
“We were not ‘wrong’ about fingerprints,” he said of forensic experts. “The unproven but intuitively true claim that no two fingerprints are ‘exactly alike’ is not rebutted by finding that fingerprints are similar. Fingerprints from different people as well as from the same person have always been known to be similar.”
According to the paper, the system could be useful in crime scenes where fingerprints were found on different fingers than those on the police record, but Cole stated that this is only possible in rare cases because when prints are taken, all ten fingers and palms are routinely recorded. “It’s not clear to me when they think law enforcement will have only some, but not all, of an individual’s fingerprints on record,” said the law enforcement official.
The study’s authors say they are confident in the findings and have open-sourced the AI code for others to review, which Champod and Cole praised. However, Guo stated that the significance of the study extends beyond fingerprints.
“This isn’t just about forensics, it’s about AI. Humans have been looking at fingerprints since we existed, but nobody ever noticed this similarity until we had our AI analyze it. That just speaks to the power of AI to automatically recognize and extract relevant features,” he said.
“I think this study is just the first domino in a huge sequence of these things. We’re going to see people using AI to discover things that were hiding in plain sight, right in front of our eyes, like our fingers.”