As you may have seen elsewhere, I’ve been wondering whether the widely-publicized Columbia University study on the uniqueness of fingerprints isn’t any more than a simple “discovery” of fingerprint patterns, which we’ve known about for years. But to prove or refute my suspicions, I had to read the study first.
My initial exposure to the Columbia study
I’ve been meaning to delve into the minutiae of the Columbia University fingerprint study ever since I initially wrote about it last Thursday.
(And yes, that’s a joke. The so-called experts say that the word “delve” is a mark of AI-generated content. And “minutiae”…well, you know.)
If you missed my previous post, “Claimed AI-detected Similarity in Fingerprints From the Same Person: Are Forensic Examiners Truly ‘Doing It Wrong’,” I discussed a widely-publicized study by a team led by Columbia University School of Engineering and Applied Science undergraduate senior Gabe Guo. Columbia Engineering itself publicized the study with the attention-grabbing headline “AI Discovers That Not Every Fingerprint Is Unique,” coupled with the sub-head “we’ve been comparing fingerprints the wrong way!”
There are three ways to react to the article:
- Gabe Guo, who freely admits that he knows nothing about forensic science, is an idiot. For decades we have known that fingerprints ARE unique, and the original forensic journals were correct in not publishing this drivel.
- The brave new world of artificial intelligence is fundamentally disproving previously sacred assumptions, and anyone who resists these assumptions is denying scientific knowledge and should go back to their caves.
- Well, let’s see what the study actually SAYS.
Until today, I hadn’t had a chance to read the study. But I wanted to do this, because a paragraph in the article that described the study got me thinking. I needed to see the study itself to confirm my suspicions.
“The AI was not using ‘minutiae,’ which are the branchings and endpoints in fingerprint ridges – the patterns used in traditional fingerprint comparison,” said Guo, who began the study as a first-year student at Columbia Engineering in 2021. “Instead, it was using something else, related to the angles and curvatures of the swirls and loops in the center of the fingerprint.”
From https://www.newswise.com/articles/ai-discovers-that-not-every-fingerprint-is-unique
Hmm. Are you thinking what I am thinking?
What were you thinking?
I’ll preface this by saying that while I have worked with fingerprints for 29 years, I am nowhere near a forensic expert. I know enough to cause trouble.
But I know who the real forensic experts are, so I’m going to refer to a page on onin.com, the site created by Ed German. German, who is talented at explaining fingerprint concepts to lay people, created a page to explain “Level 1, 2 and 3 Details.” (It also explains ACE-V, for people interested in that term.)
Here are German’s quick explanations of Level 1, 2, and 3 detail. These are illustrated at the original page, but I’m just putting the textual definitions here.
- “Level 1 includes the general ridge flow and pattern configuration. Level 1 detail is not sufficient for individualization, but can be used for exclusion. Level 1 detail may include information enabling orientation, core and delta location, and distinction of finger versus palm.”
- “Level 2 detail includes formations, defined as a ridge ending, bifurcation, dot, or combinations thereof. The relationship of Level 2 detail enables individualization.”
- “Level 3 detail includes all dimensional attributes of a ridge, such as ridge path deviation, width, shape, pores, edge contour, incipient ridges, breaks, creases, scars and other permanent details.”
We’re not going to get into Level 3 in this post. But if you look at German’s summary of Level 2, you’ll see that he is discussing the aforementioned MINUTIAE (which, according to German, “enables individualization”). And if you look at German’s summary of Level 1, he’s discussing RIDGE FLOW, or perhaps “the angles and curvatures of the swirls and loops in the center of the fingerprint” (which, according to German, “is not sufficient for individualization”).
Did Gabe Guo simply “discover” fingerprint patterns? On a separate onin.com page, common fingerprint patterns are cited (arch, loop, whorl). Is this the same thing that Guo (who possibly has never heard of loops and whorls in his life) is talking about?
I needed to read the original study to see what Guo actually said, and to determine if AI discovered something novel beyond what forensic scientists consider the information “in the center of the fingerprint.”
So let’s look at the study
I finally took the time to read the study, “Unveiling intra-person fingerprint similarity via deep contrastive learning,” as published in Science Advances on January 12. While there is a lot to read here, I’m going to skip to Guo et al’s description of the fingerprint comparison method used by AI. Central to this comparison is the concept of a “feature map.”
Figure 2A shows that all the feature maps exhibit a statistically significant ability to distinguish between pairs of distinct fingerprints from the same person and different people. However, some are clearly better than others. In general, the more fingerprint-like a feature map looks, the more strongly it shows the similarity. We highlight that the binarized images performed almost as well as the original images, meaning that the similarity is due mostly to inherent ridge patterns, rather than spurious characteristics (e.g., image brightness, image background noise, and pressure applied by the user when providing the sample). Furthermore, it is very interesting that ridge orientation maps perform almost as well as the binarized and original images—this suggests that most of the cross-finger similarity can actually be explained by ridge orientation.
From https://www.science.org/doi/10.1126/sciadv.adi0329.
(The implied reversal from the forensic order of things is interesting. Specifically, ridge orientation, which yields a bunch of rich data, is considered more authoritative than mere minutiae points, which are just teeny little dots that don’t look like a fingerprint. Forensic examiners consider the minutiae more authoritative than the ridge detail.)
Based upon the initial findings, Guo et al delved deeper. (Sorry, couldn’t help myself.) Specifically, they interrogated the feature maps.
We observe a trend in the filter visualizations going from the beginning to the end of the network: filters in earlier layers exhibit simpler ridge/minutia patterns, the middle layers show more complex multidirectional patterns, and filters in the last layer display high-level patterns that look much like fingerprints—this increasing complexity is expected of deep neural networks that process images. Furthermore, the ridge patterns in the filter visualizations are all generally the same shade of gray, meaning that we can rule out image brightness as a source of similarity. Overall, each of these visualizations resembles recognizable parts of fingerprint patterns (rather than random noise or background patterns), bolstering our confidence that the similarity learned by our deep models is due to genuine fingerprint patterns, and not spurious similarities.
From https://www.science.org/doi/10.1126/sciadv.adi0329.
So what’s the conclusion?
(W)e show above 99.99% confidence that fingerprints from different fingers of the same person share very strong similarities.
From https://www.science.org/doi/10.1126/sciadv.adi0329.
And what are Guo et al’s derived ramifications? I’ll skip to the most eye-opening one, related to digital authentication.
In addition, our work can be useful in digital authentication scenarios. Using our fingerprint processing pipeline, a person can enroll into their device’s fingerprint scanner with one finger (e.g., left index) and unlock it with any other finger (e.g., right pinky). This increases convenience, and it is also useful in scenarios where the original finger a person enrolled with becomes temporarily or permanently unreadable (e.g., occluded by bandages or dirt, ridge patterns have been rubbed off due to traumatic event), as they can still access their device with their other fingers.
From https://www.science.org/doi/10.1126/sciadv.adi0329.
However, the researchers caution that (as any good researcher would say when angling for funds) more research is needed. Their biggest concern was the small sample size they used in their experiments (60,000 prints), coupled with the fact that the prints were full and not partial fingerprints.
What is unanswered?
So let’s assume that the study shows a strong similarity between the ridges of fingerprints from the same person. Is this enough to show:
- that the prints from two fingers on the same person ARE THE SAME, and
- that the prints from two fingers on the same person are more alike than a print from ANY OTHER PERSON?
Or to use a specific example, if we have Mike French’s fingers 2 (right index) and 7 (left index), are those demonstrably from the same person, while my own finger 2 is demonstrably NOT from Mike French?
And what happens if my finger 2 has the same ridge pattern as French’s finger 2, yet is different from French’s finger 7? Does that mean that my finger 2 and French’s finger 2 are from the same person?
If this happens, then the digital authentication example above wouldn’t work, because I could use my finger 2 to get access to French’s data.
This could get messy.
More research IS needed, and here’s what it should be
If you have an innovative idea for a way to build an automobile, is it best to never talk to an existing automobile expert at all?
Same with fingerprints. Don’t just leave the study with the AI folks. Bring the forensic people on board.
And the doctors also.
Initiate a conversation between the people who found this new AI technique, the forensic people who have used similar techniques to classify prints as arches, loops, whorls, etc., and the medical people who understand how the ridges are formed in the womb in the first place.
If you get all the involved parties in one room, then perhaps they can work together to decide whether the technique can truly be used to identify people.
I don’t expect that this discussion will settle once and for all whether every fingerprint is unique. At least not to the satisfaction of scientists.
But bringing the parties together is better than not listening to critical stakeholders at all.
Bredemarket 