Fingerprints have been used in criminal investigations for more than a hundred years. Until now, it was assumed that every fingerprint is unique and that the prints from different fingers of the same person are no more similar than those of any other person. Researchers have now refuted this assumption with the help of artificial intelligence. A suitably trained AI model can determine with a hit rate of 77 percent whether the prints of two different fingers come from the same person or not. This could expand the possibilities of forensics in the future.
Fingerprints can help convict criminals. To date, however, it has been necessary that the prints secured at a crime scene come from the same fingers as those previously stored in a database. If, on the other hand, the print at the crime scene comes from a middle finger, for example, but only prints from the index finger are stored, no classification has been possible so far. Even though prints from different fingers of the same person were found at different crime scenes, no connection could be made. According to the current dogma of forensic science, every fingerprint is unique, even if it comes from different fingers of the same person.
Similarities revealed
A team led by Gabe Guo from Columbia University in New York is now questioning this assumption. “Contrary to prevailing belief, we show with over 99.99 percent confidence that fingerprints from different fingers of the same person show very strong similarities,” the team reports. Using 60,000 fingerprints from a public US government database, the researchers trained an artificial intelligence to recognize whether two fingerprints came from the same person or not. The AI model achieved a hit rate of 77 percent.
But how can AI assign the prints of different fingers to a person with a relatively high degree of reliability if there are supposedly no similarities? To answer this question, the research team analyzed in detail which clues the AI relies on. In conventional fingerprint analyzes - be it in criminalistics or for biometric scanners - a system evaluates the so-called minutiae, i.e. the end points and branches of the fine grooves of the fingerprint. “Our AI does not use minutiae, but instead evaluates the angles and curvatures of the swirls and loops in the center of the fingerprint,” explains Guo.
Not yet for real use
For further analysis, the team subgrouped test subjects' fingerprints by gender and ethnicity and tested whether the similarities between a person's fingers increased or decreased depending on gender or ancestry. However, the model was most reliable when trained with fingerprints from all demographic groups. “This shows that the similarity is highly generalizable,” the researchers conclude.
They hope that AI-supported methods can help to better identify criminals or exonerate innocent people in the future. “Our experiments suggest that in some situations this relationship can increase the efficiency of forensic investigation by almost two orders of magnitude,” the team said. However, the current system is still too imprecise for real use in law enforcement. “However, we assume that the proposed system will achieve significantly better performance when trained with very large government databases that also contain partial fingerprints.”
New type of science thanks to AI
According to the authors, the study also shows how artificial intelligence can change science. None of the authors had any previous contact with forensics. Instead, they are engineers and computer scientists. Without having previously studied fingerprints in detail, Guo began the project as an experiment during the first semester of his studies, based on a public data set combined with a relatively simple AI model.
“The current paper shows that a student with no forensic background can use artificial intelligence to successfully challenge a widely held assumption of an entire field,” says one of his supervisors, co-author Hod Lipson. “We are facing an explosion of AI-powered scientific discoveries by non-experts, and the expert community, including academia, must adapt to this
Source: Gabe Guo (Columbia University, New York) et al., Science Advances, doi: 10.1126/sciadv.adi0329