AI flips the script on fingerprint lore – maybe they're not so unique after all The belief that all fingerprints are unique is so well accepted that crime novels and TV shows riff on it. Recent AI research has challenged this notion, at least regarding the fingerprints on different fingers of the same person. Undergrad researchers at Columbia Engineering found that while the branching and endpoints in the …
So basically, having multiple partial prints from different fingers can now be narrowed down to a smaller number of individuals than could be done by previous methods. Matching a know individual to a partial print still requires manual comparison for proof.
I'm not surprised that all my fingers have an underlying similarity at this level, they all derived from the same DNA at the same time.
Matching a know individual to a partial print still requires manual comparison for proof.
I think you mean "proof", with scare quotes. Studies have shown wide discrepancies among experts in "matching" fingerprints.
Fingerprint identification is one of the great myths of criminal forensics. Sure, many times it's right, because the odds are pretty good when you have additional evidence to update on. But (as the article notes) the forensics industry has long made hugely overblown and unsupported claims about its accuracy and precision.
I read an article years ago form someone who'd done extensive research into the rhetoric from forensic scientists and law enforcement agents around fingerprint identification. (He was called as an expert witness by the defense in a criminal trial, but the judge refused to allow his testimony on the grounds that "fingerprint identification is established science".) It's full of unscientific claims, unsupported assertions, and logical fallacies.
It's not quite the same level of patent bullshit as, say, facial reconstruction, to say nothing of outright scams like bite identification; but it's not nearly the magic wand that the police and Hollywood would have us believe.
Only very minor ones. The primary objective in forensic fingerprint anaysis is to identify a suspect person -- not to identify which finger they used in the crime. It looks like the uniqueness of fingerprints for forensics remains valid for the time being.
Hmm.
77%. 50% would be random guessing. 100% would be certainty. 77% isn't very good at all, without some better stats behind it. 23% wrong, it sounds like something they'd have used for the Birmingham 6! 1:3, that's good odds for a mistrial.
Not so.
Let's say you have have prints at both scenes that all read as different people. Now maybe you can link a print at each scene together as the same person, making the person matching those prints as higher interest.
Now combine that with other evidence. Let's say you don't have any reason to believe two crimes are linked, and had cell location data showing hundreds of people present at each scene in the time in question. If you were able to establish that one person was at both scenes, then you can consider only the people who cell data shows were at both scenes to try to figure out who that person is. Are they are a victim? Are they are a suspect?
looks like the uniqueness of fingerprints for forensics remains valid for the time being.
It is well known that fingerprints are not unique. Like DNA, the only guarantee is that fingerprint characteristics are unusual -- there aren't many people out there who will be identified as having the same DNA signature or fingerprint signature. The fingerprint signature depends on having a complete or nearly complete set: since individual fingerprints are far from unique, you need a set of 10 to make a reliable fingerprint signature. Of course hash conflicts do happen -- their have been a few people out there who are unfortunate enough to have fingerprint signatures matching that of wanted criminals.
Given that "It is well known that fingerprints are not unique.", it is rather unfortunate that this Register article manages to attribute the exact opposite statement to some unknown journal rejection.
FWIW, the new information is about the similarity of different fingers on one person. It (sort of, in an unreliable way) manages to help build up a complete set of fingerprints from individual fingers of that set. Having a set enables you to build the set signature that enables you to do fingerprint identification.
The prints were incomplete and a "possible match" for about 20 people in the FBIs database. The Spanish authorities declared that there was no match with the crime scene. Later, a US judge declared that the evidence was largely fabricated.
My memory is not as reliable as it was, but I recall that in the 1970s it was determined that for a full match for 2 different individuals from a single print was about 1 in 2 billion, or roughly 2 possible matches for the World's population at that time.
Figures which have been largely debunked
Fingerprints developed from having a small set of known criminals in a small geographic area - where a match was more than likely to be accurate simply because of the statistically small number of the population engaged in crime
Once it started casting a wider net and began using computerised "point matching" it was quickly realised that there were insufficient points being used to separate unique (but similar) fingerprints
A "full match" is all 10 fingers, but crime scene forensics are lucky to get more than a couple of partials most of the time. When you have that, your hit rate drops dramatically and ISTR reading the odds of such matches could be as low as 1 in 150,000 if ALL the population is fingerprinted (again, there's a self-selection part to this, where for the most part what's on file is mostly known criminal and their associates) and there has been at least one case of someone matching fingerprint evidence in a case dating from before they were born
"In order to understand that it was merely identifying the angles and starting points of the ridges, they had to study the AI system's decision process. Thus, the team concluded that the AI was using an unexpected forensic marker."
What is interesting to me is that the design of their pattern-matching system (not AI, but I get that real researchers sadly need to use this term to get any funding) allowed them to identify WHY the results being given were being given, which I believe is (by design or because of huge complexity) not possible in the more complex LLMs.
It IS very surprising to find that part of a single individuals' prints could be identical across fingers.... not sure exactly what the forensic utility is though, since normally you want to use fingerprints to distinguish between different *people* not different *fingers*. I suspect that rather like DNA, forensic comparison uses a simplified mathematical model of the fingerprint or DNA to be able to more easily compare them, and model matches might not necessarily be aligned with real-life matches. So in any case it's good that some "thought-to-be-correct-but-not-really-proven" assumptions such as unique fingerprints are being challenged.
identify WHY the results being given were being given, which I believe is (by design or because of huge complexity) not possible in the more complex LLMs
The situation is more complicated than that (when isn't it?). There's a lot of research into explicability for SotA LLMs and other large models, and some of it has produced interesting and useful, if far from complete, results. (Here's a survey from a while back.) Conversely, we're seeing more research like that described in the article, where teams use models that already are explicable; that's helping us scale those explication techniques up to larger and more-complex models. Those two trends are approaching the problem from the opposite ends, as it were, and may yet converge.
But yes, you're right that using explicable models is the Right Way to do computational science with deep ANN stacks like this. The recent1 paper in Nature from Wong et al., where the team used an explicable deep-learning system to identify a novel class of antibiotics, is another good example.
1But submitted two years ago, in January 2022.
the design of their pattern-matching system ... allowed them to identify WHY the results being given were being given
You're assuming something without any basis. If they have fingerprints matching and between fingers on the same person and they can't understand how it is doing that, it doesn't follow that their system allowed them to get inside the model and figure out how it was doing that.
Maybe they said "wow, that's weird, how is it doing that?" and they figured it out by training it on only portions of fingerprints and see what matched, until they figured out how it was doing so.
"Many people think that AI cannot really make new discoveries – that it just regurgitates knowledge," said Lipson. "But this research is an example of how even a fairly simple AI, given a fairly plain dataset that the research community has had lying around for years, can provide insights that have eluded experts for decades."
AI can't make new discoveries. It just shows how poor these "experts" were if they couldn't see that.
These are the dangers of "science is settled" approach, many so called $cientists take.
> It's like saying microscope made new discoveries and not the person using it.
I'm not saying that you're completely wrong, but that's a poor comparison. A microscope is completely inert.
Certainly "AI making a new discovery" is a bit of a stretch and overstating the situation. Better to say that this tool revealed to us a novel aspect.
Oh, please. Ken Wilson (who had, y'know, a Nobel in Physics, which relatively few Reg commentators hold) started talking about computation as the "third leg" of science in, what, the mid-1980s? Using "the machine discovered" as synecdoche1 for "a whole bunch of humans, using a vast array of technology, in a wildly complex social arrangement, and drawing on many generations of intellectual and physical labor, made a discovery" is hardly novel.
Bitching about "AI making a discovery" is a cloaked appeal to the Romantic idea of individual genius. The nineteenth century is 124 years back thataway, guys.
1Pale Moon's English dictionary doesn't include "synecdoche"? Honestly, does no one command a decent vocabulary these days?
AI can't make new discoveries
You're wrong, at least for a suitably flexible definition of "AI" (which is a largely meaningless term anyway).
Experts brought into court to compare fingerprints only compare a small set of features. By simplifying the way prints are compared, finding a match is also greatly facilitated. People have been convicted on the basis of a small part of a fingerprint, when the expert stands before the jury and solemnly declares that this small trapezoid matches that one, demonstrating, in the expert's experience, that the committer of the crime and the possessor of the finger are one and the same person.
Correct, and also a direct followon to the ICL Pathway debacle - Post Office was under tremendous pressure to prove its previous £1.5billion software flop(*) had been remediated
(*) The largest corporate IT failure in history at the time (1999) and Horizons is built directly on the ruins of that project
DNA evidence is also quite bad these days, the samples are analysed by unaudited proprietary algorithms and programs, and there have been multiple cases of one program saying it is a match and another saying it is not a match.
The development of this sort of software should really be government funded and open source.
There was the world leading Forensic Science Service. Until it was cut in 2012, when the tories realised they could just get rid, saving all those appeals and "awkward" not guilty verdicts!
https://en.m.wikipedia.org/wiki/Forensic_Science_Service
SEE: DNA evidence miscarriages of justice
Whilst DNA has cleared a lot of people, there are a small number where "expert testimony" on DNA has put the wrong people in jail
The base problem is that you're using a good tool for _eliminating_ suspects as an unquestioned one for matching them. It should (must) be only part of the evidence, not the entire basis of the case (ie: if a match is obtained, it's only an initial filter for closer inspection at higher resolution)
This is more how I understood fingerprint evidence to work: That it eliminates those who do not match, leaving those who do, which is normally just one person.
The original finger print argument was (as I recall) that if you took 100 prints, then took any 1 of those 100 and compared them to the rest, you didn't get a match with any of the others. In this regard, the fingerprints are sufficiently different that they can be considered 'unique'. However, if you took prints from someone today, and compared them to 100,000 prints taken 10 years ago, there's a higher probability you'll get a false positive rather than a true positive: That the new prints will match sufficiently to someone else's prints, but not to the older copy of their own. This was due to the prints migrating/changing over time.
Bottom line was: For the purpose of a limited group shortly after a crime was committed and evidence obtained, fingerprints were good enough to narrow down the list of suspects to one person. Beyond that, the trust in fingerprint matching depreciates towards zero.
The above has been from articles both for and against fingerprints as evidence, particularly where research has shown support for the claims both for and against the reliability of fingerprints.
Bringing 'AI' into the mix is the same as introducing any fast processing system: It does the work faster, and potentially can go into finer detail. It doesn't make the process 'new', however: It remains pattern recognition and probability/predictive modelling.
There are huge databases of fingerprints in existence, for example for visiters to the US and soon for British visitors to EU countries. Can you be sure that fingerprints will be used only to eliminate suspects* and not to find them? After the next major terrorist attack** all possible means will be used to identify suspects and the pressure to get convictions will mean that doubts could well be swept aside.
*from the enquiry.
** surely a when not an if.
Yes, but it's fairly easy to look at old auntie Mavis, 350 miles away in a care home, and swiftly eliminate her from the list of 4 suspects.
It is very rare that anyone would try to convict on a single fingerprint. You'd be either really unlucky, or it's the post office.
Unfortunately it's not rare and fingerprint evidence credibility has been built up time and time again by "expert testomony" in front of juries who don't understand statistics and have only a passing acquaintance with scientific rigour
Yup, the old base-rate effect comes into play again.
If there the odds of a false match are a million-to-one against, and you are comparing the prints taken from a scene against those of a few dozen suspects who could have been around the scene around the time of the crime, then you can probably place a high level of confidence in a match. But if you are comparing them against a database of an entire country’s population of, say, five million (make it a small country), then you will likely have 5 false positives.
This is more how I understood fingerprint evidence to work: That it eliminates those who do not match, leaving those who do, which is normally just one person.
That may well be how fingerprint evidence works. It's not how, as a matter of practice, it is used, by law enforcement and the courts.
The police are incentivized to close cases. Prosecutors, at least under an adversarial system such as that of the US,1 are incentivized to win them. Forensics labs are incentivized to provide "identifications" and evidence of guilt, since they're primarily funded either by the state or by law enforcement.
None of that encourages scientific rigor.
1And all too often even in systems which are in theory not adversarial.
> That it eliminates those who do not match, leaving those who do, which is normally just one person.
That may be the procedure, but not sure it addresses my original question: my guess would be that matching prints, especially if they're partial, is not necessarily, or even in general, all-or-nothing, and might best be expressed as a probability. That probability might be (at least roughly) assessed by running statistics on large random samples of prints (adjusting for partialness, demographics, and other relevant factors), of which there is hardly a shortage.
Interesting article: https://thecrimereport.org/2021/12/06/is-a-single-fingerprint-enough-to-convict/
From the article: While fingerprint matching, for example, is now a common tool in law enforcement investigations, there is still a lack of sufficient empirical studies that conclusively establish the uniqueness of the fingerprints, according to a study published in South Africa’s Potchefstroom Electronic Law Journal.
1:2800000000 is beyond reasonable doubt, when you're dealing with a country with numbers of people in the millions in it at any one time.
Divide the big number with how fractionally complete the collected fingerprint sample is. Then look at all the other evidence that's not fingerprints.
What any jury is doing is Bayesian analysis, they just generally don't know that's what it's called. As for reasonable doubt, some jurors have different amounts of reasonableness. That's why there are usually 12 of them.
Fingerprints became an established law-enforcement tool before statistical analyses had really caught up. Whereas DNA matching, for comparison, came along after such statistical analysis techniques were already well-known. This has resulted in DNA matching being held to quite different standards, more rigorous than fingerprint matching.
Doubts have been raised over the uniqueness of fingerprints before. But of course giving any credence to those doubts, after so many decades of using them as crucial evidence in court cases, would be seen by many in law enforcement as rocking the boat just a little too much. Which is why you see such reluctance to admit them.
the angles and curvature at the center of the fingerprint could be the same across an individual
How large an area are we talking about, because I'm looking at my index fingers and they're completely different. One does the usual swirly thing. The other sort of swirls at the tip end but the other goes down in straight lines until it suddenly veers off to the side.
insights that have eluded experts for decades
Was it really eluding the experts, or were the experts simply peddling the "they're all unique" line? I think uniqueness depends upon the accuracy of your equipment and what you're checking...a bit like the uniqueness of DNA testing.
You're in the 33% of people it really doesn't work for, I suspect.
Looking at my fingers, excluding thumb, I've long noted that they're almost identical on one hand, just scaled up and down. Who hasn't, when doing the kids activity of fingerprinting yourself as a "detective" with a cheap kit of graphite, ink pad and talc? (You, because they aren't)
Looking at mine, they're nearly identical (or at least highly similar) on all fingers (including thumbs) for BOTH hands, with the oddity that the index fingers on both hands are mirrored to the rest of the fingers on that hand and the hands themselves are mirrored
Fingerprints (like fur patterns on cats) are highly influenced by in-utero conditions and it's not at all unusual for identical twins to have substantially different fingerprints (although likely exhibiting the same basic characteristics)
Not sure how this changes things apart from potentially increasing the probability of an id hit on an individual when there is more than one partial print available.
They have not said that prints are matched between individuals, but a sample size of 60,000 is tiny in this context, assuming standard practice of having five appendages and two hands that's only 6000 individuals ... Be interesting to see what happens if it is run against a big database to see if the pattern matching happens between more individuals. If it turns out that 1 in 10000 are matched, how does that impact security of conviction based on fingerprint evidence alone (if that's a thing)?
Fingerprinting needs to be put in context with how policing works (or did) and the composition of the persistent criminal community - figures are from a few years ago but I doubt that they have changed that much.
For example in the UK over 75% of all crime is committed by around 100,000 people.
For any given crime, 4 times out 5, the perpetrators will already be known to the police - hence the "usual suspects".
For the missing 25% a large proportion is youth crime performed by the up and coming - and consequently the reason for the emphasis to cut youth crime.
Already around 20 - 25% of crime is committed by those under 21 years of age.
Once you know this, you can see how finger printing helps narrow down to persons of interest with generally very good results as your target sample size is neither huge nor random.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
