An Impenetrable Program Transforming How Courts Treat DNA Evidence | Hacker News Reader

An Impenetrable Program Transforming How Courts Treat DNA Evidence

wired.com

46 pointsrbanffy18 comments

This seems to me a bit like a zero-knowledge proof. The company wants to hide it's source code, but claims the source code can identify DNA evidence correctly. The public wants to believe them, but how can we?

Well, what if every time the prosecutor wanted to use such evidence, they had to include say 10 other samples of to test and the DNA of 10 other suspects unrelated to the case? It would be more expensive I imagine, but it could lead to more fair evidence- if this technique really does work, it should be able to accurately identify the noise variables and point out the signal. If there is no signal, great. If there's lots of false signals, then we know this is bullshit.

Truthfully, my worry here is that if you have 5 DNA testing companies, and 1 always gives the prosecutors the answers they want, they'll be the ones who get more business. We need a way to ensure accuracy.

Source code shouldn't matter for performance testing though.

Simply fire through a bunch of known-result-samples through the various services and take a look at the test statistics. Determining accuracy is hardly the hard part.

We cant test all edge cases as well as we can read source. Handicapping justice for a dollar is whoring of the lowest order.

We can't find all the edge cases from reading source as well. But we also don't need to test all edge cases, it's still about probability, if the lab was tested (some double-blind study) and could identify successfully 1000 cases out of 1000, it means that probability of some random people being an edge cases where algorithm is not as accurate is less than 0.1%. We just need to make sure such tests/certification is used and there's no ties between labs and prosecutors. Of course it probably would be nice the company to show their process openly, but you still will need these tests.

Double-blind tests are necessary but not sufficient. Without the double-blind tests, the process simply has no credibility at all. The double-blind tests establish that credibility.

However, the 6th amendment also guarantees defendants the right to confront their accusers. Without the source code, they are unable to do this.

How have courts obtained the source code of human eyewitnesses all these years?

Obviously there is no source code, but you can cross-examine a human eyewitness. There is no way to cross-examine an algorithm. Eyewitnesses that can't be cross-examined are only allowed in court in very rare circumstances, such as when the defendant killed them to prevent them from testifying.

> “It would take a genius to work out an error from a code.”

Really? Better keep the algorithm public then where many “geniuses” have access to it, and can scrutinise it, considering people are being put away with “multiple lifetime” sentences based on evidence generated by it.

Outrageous that this is ever allowed in court. A trial is supposed to involve the prosecution publicly proving that the accused person is guilty, not to simply have guilt handed down by some inscrutable black-box.

Violation of the confrontation clause.

chad@flyingdogsolutions.com if any attorney needs an expert witness to look into it. I can review the source under a protective order so the company's IP will be fully protected. Worst case I find something wrong and they fix it.

> As John Oliver commented last month...

That pretty much says it all.

This is like requiring the manufacturer of glassware used in the labs to hand over their proprietary manufacturing information, or all the evidence processed by the lab must be thrown out. The glassware can be verified to work properly given appropriate testing, etc. So can the software without looking at the source.

In fact, treating the software as a black box and employing appropriate testing without consideration as to what's in the box - just what the box should be able to do - should lead to more confidence.

The real issue here is whether there is such a certification / testing process that the tool meets that can be independently verified.

> In October, the Department of Commerce National Institute of Standards and Technology announced that it would embark on a study to determine the reliability of DNA testing, including the algorithmic methods used by companies like TrueAllele. This research, which NIST says is to establish “foundational validity,” has been condemned by Perlin specifically through a post on the Cybergenetics blog, which calls the study “wasteful, “unnecessary,” and “meaningless.” Perlin argues that True Allele’s validity has already been proven through scientific, peer-reviewed research studies. But these studies that Perlin relies on are internal validation studies that were paid for and run by Cybergenetics.

It is a bit unsettling that NIST wants to independently test “foundational validity” but the company founder is so dismissive of the idea.

Since STRmix, a competing algorithm, is already public, I am curious whether publication led to new defense. Such case would be a strong evidence for the proposition that "the source code is crucial to defense".

Well this is all stupid. Because we have double blind testing. It doesn't matter what the algorithm is, it just matters whether the test is accurate. And we can determine that by sending in test samples and having them rate these. If we want to test its accuracy when working from the minimal DNA included in a fingerprint then have a few people touch a spot and send in a swab to compare (sometimes ask to compare to the DNA of the person who touched it, sometimes to a different person).

A blind study means don't tell the DNA testing company which submissions are for real cases and which are for testing. A double-blid study means don't tell the police submitting it either (to avoid giving subtle clues to the testing company, like calling to follow up on tests but not real cases, or vice versa).

If they (the testing company and the police) are willing to conduct a double-bind study, then its results are a reliable way to assess the technology. If they are NOT willing to conduct a double-blind study then we should ask why not. The answer is almost certainly because they are afraid of what the results might show.

> Well this is all stupid. Because we have double blind testing. It doesn't matter what the algorithm is, it just matters whether the test is accurate.

Seems to me that your kind of reasoning is why the Volkswagon cheating went undetected for so long.

What does this have to do with Volkswagen? A double-blind study using that test would have shown that the test was faulty for the model of cars that were cheating.

It went undetected so long because the regulators were not competent and constructed a test that was easy to game.

Science is best based on techniques such as properly formulated studies than source code analysis. Consider how many bugs fuzzing has uncovered that escaped source code analysis for many years by many experts.

Every step of the judicial process should be completely open to scrutiny in every particular by the accused and in general (eg exact methods) by any motivated citizen. This transparency is more important to the long-term viability of our legal system than extra accuracy provided by a proprietary algorithm. It is possible that this analysis is not amenable to monetization at the level necessary to sustain private investment, which is a different problem.

Reply on news.ycombinator.com