AI fingerprint breakthrough could aid future forensics

The field of forensics has long assumed that no two fingerprints are ever alike, even on different fingers from the same person. But, analysis from scientists at the Creative Machines Lab at Columbia University challenges this conventional wisdom. By asking an artificial intelligence to take a look at a suite of tens of thousands of publicly available fingerprints, Gabe Guo has found that there are features that are more similar in prints from the same person, after all...

Gabe - This research started as a chat between me and Professor Wenyao Xu at the University of Buffalo. And me and him were just having a chat and he asked me at the end, 'do you think that all fingerprints are really unique? Think about fingerprints from different fingers of the same person. They come from the same genetic material, so don't you think they might be correlated?' And once he said that, I was like, that's a really interesting idea. Let me go investigate this. And to bring a long story short, little did I know that one conversation would form the basis of my life's focus for the next three years

Chris - Are all fingerprints of a person not in some way similar then, because I always thought that fingerprints from an individual had common features or factors, which is how we identified a criminal for example.

Gabe - I think there may have been a hunch among some people that there is a relation between different fingerprints from the same person, but no one was ever able to provide quantitative or explainable evidence to substantiate this claim. And that's really our contribution. We were the first in the world to systematically analyse it, use statistics and say there is this similarity and specifically we can get more into that later. It comes from the angles of the fingerprint ridges at the centre.

Chris - Does that mean then that when the police go and investigate, say a crime scene or border control are collecting people's fingerprints, that if they had two fingerprints from different fingers, they previously would've said, 'well we can't say that that was the same person.'

Gabe - Right. They would have no way of matching them before our discovery.

Chris - So what have you done then that's turned the tables on all this?

Gabe - We wanted to come at this with a unique perspective because we knew that forensic science was not able to correct the question. So we thought, okay, what about we just feed it to an AI model? And really the power behind AI models is they can automatically learn what the relevant parts and features and characteristics of an image are that they should focus on. So we fed our fingerprints into an AI model. The idea was just to say, 'okay, are these two fingerprints from the same person or from a different person?' And after analysing tens of thousands of fingerprints, the AI model was able to not only identify them with very high accuracy, but also after looking into what the AI model was analysing, we were able to find the causes of the similarity, which is the angles of the fingerprint ridges at the centre.

Chris - Most AI models though are not what we call explainable. When you ask them how they reach the conclusion that they do, they can't tell you. And this is one of the reasons why people tend to not trust them or have some degree of scepticism. So how does yours explain to you what it's seeing and to show that two sets of prints from the same person have these relationships?

Gabe - Yeah, that's a great question. And actually that was the sticking point and the thing that took us the most time with the research. So there are two ways to do it and we employed both ways. The first way is what's known as ablation. You take the data you have and you degrade it in some way. So for our data we said okay, if we make it only black and white, can it still find the similarities? Okay, that works. What if we only have the angles of the fingerprint ridges and we take out all notion of thickness or colour. And that still works to almost the same accuracy. So we were able to say, okay, we're pretty sure that the similarity comes from the angles of the fingerprint ridges regardless of colour or thickness. The second way we analysed it was we saw which parts of the images caused the AI's neurons to light up the most. And after systematically doing that a bunch of times we found out that almost always it was the centres of the fingerprints that caused the AI's neurons to really fire.

Chris - So explain that for us a bit then. What exactly is it that it's found that shows these similarities? I know you said there is a region, but what is it about those bits of the fingerprint that are so similar between individuals?

Gabe - Fingerprints fall into some categories, some fingerprints loop back in on themselves. There's some fingerprints that don't really loop back in on themselves and so on and so forth. It looks at that same region, but it doesn't exactly look at just that macro characteristic. It looks at something a little bit more subtle. The angles in that central region where you kind of have the curvature determined, but then there's some subtle things it's doing that kind of mould all those together into the final prediction.

Chris - And how good is it in the sense that how hard would I have to try in percentage terms to fool it? How often does it get it wrong? Did you find any situations where it would slip up? Because of course that is critical. When it comes to forensics, a jury wants to be X number of decimal places percent sure that this is not happening by chance.

Gabe - So as currently constructed, this technology cannot be used to be deciding evidence in the court case because depending on the conditions, the accuracy ranges from 75% to around 90%. So as you mentioned, it's still not good enough to be used as evidence in court cases.

Chris - So do you think you can improve on that or is this just a useful tool it gives police or investigators the reassurance that they need sometimes that they're on the right lines, they need to go and find other corroborating evidence? Or do you think you can get this up to the magic sort of 99.99% that it's probably going to take to be admissible?

Gabe - Yeah, I believe that we can get it to that high state of 99% because when we did studies investigating how much data our model needed, we found a very strong trend that, as you add more data, the performance shoots up a lot. And unfortunately for this study we were limited by there being only 60,000 publicly available fingerprints. But I'm sure that after this study comes out, we'll be able to get access to more fingerprints and train a much stronger model.