Slashdot Mirror
Anti-Terrorist Data Mining Doesn't Work Very Well
Presto Vivace and others sent us this CNet report on a just-released NRC report coming to the conclusion, which will surprise no one here, that data mining doesn't work very well. It's all those darn false positives. The submitter adds, "Any chance we could go back to probable cause?" "A report scheduled to be released on Tuesday by the National Research Council, which has been years in the making, concludes that automated identification of terrorists through data mining or any other mechanism 'is neither feasible as an objective nor desirable as a goal of technology development efforts.' Inevitable false positives will result in 'ordinary, law-abiding citizens and businesses' being incorrectly flagged as suspects. The whopping 352-page report, called 'Protecting Individual Privacy in the Struggle Against Terrorists,' amounts to [be] at least a partial repudiation of the Defense Department's controversial data-mining program called Total Information Awareness, which was limited by Congress in 2003."
I bet this will not change what they are doing or how they are doing it one bit.
Prediction: The real iPhone killer is going to be sex robots from Japan. Think about it.
The Constitution is there for a reason.
you had me at #!
In other news, water is wet, the Pope is Catholic, and Ursines excrete solid wastes in silviculture.
That is all.
And several billion dollars.
And unrestrained access to all of the personal information about everyone that can be gotten by whatever means.
It'll probably still suck then, too.
What we really need are spies. Not so much in the US, here good old fashioned detective work (with Warrants) should work.
But over seas a standing army isn't going to do anything to quell terrorism. Tanks and plans will only inspire more terrorism. What we need are good old fashioned black ops. Undercover agents penetrating the terrorist groups and talking to the bad guys. Much less collateral damage as well.
We'd get a lot further with a couple guys with silenced pistols rather than a whole army.
Eschew Obfuscation
As any Cold War spy can tell you, if you "fit the profile" of a normal law-abiding person with just enough "off-perfect" things in your life so you don't seem "too perfect," it's much easier to blend in.
Knowledge is how to play a game, intelligence is how to win, wisdom is knowing what game to play.
enough data in any kind of real time to make this work.
Years ago, we were playing with a design of a system to track all the phone calls made on the AT&T network over a 3 month period. (not record the calls, just track the billing info). The machine that management wanted to try and do it on could not hold enough data just to store the data, let alone process it. And that was the largest theoretical model of hte machine there was (about 4 times the size fo the largest one in use at the time). They really needed one about 10 times as large as the largest theoretical one, just to store the data!
Multiple that by the rest of the items one buys during the day, and we can not track all the daa that is out there.
Why did they even waste the money to do the testing and the reports?
I thought we already knew this. If the algorithm comes back with even .1% false positives the system is totally worthless. There's 365 million people in the US, .1% means that the FBI/CIA/NSA would have 365,000 people to investigate. Now go and talk to someone in the AI field and see if even .1% false positive is possible.
I'm betting that if a system is going to catch any decent percentage of terrorists (greater than 50%) the false positive rate will be above 1%. Even if you only apply the system to a relatively small number of people (say people entering a leaving the country) you are going to have hundreds of thousands, if not millions, of people to investigate. Combine any kind of realistic false positive rate with the fact that about .00001% of the population deserves to be investigated and the system is worse than worthless; all it will do is distract from the people who should be investigated.
Vote for me.
I'd take their "no fly" list and identify every single person on it who was a legitimate threat and either have them under 24 hour surveillance or arrested.
The mere concept of a list of names of people who are too "dangerous" to let fly ... but not dangerous enough to track ... that just fucking stupid.
Think about how many people could be killed in the airport terminal itself WITHOUT getting on a plane ... say during the Thanksgiving or Christmas rushes there.
What idiot would let the people on that list (if they were really a threat) into a terminal? Wouldn't you expect them to STOP them BEFORE they get into a position to do that kind of damage?
As can be seen in the recent "terrorist" arrests in the US. Once you start paying people to turn in "terrorists", you start a market in "terrorists".
So the guy who wants to sell a "terrorist" to the government finds some idiot who meets the basic criteria (non-Christian, non-white) and encourages that idiot to make inflammatory statements while being recorded.
Ka-CHING!
I know this is slashdot and all, but if anybody's actually interested in looking at the full report, it's available for reading in pdf format online.
The biggest problem is actually not the false positives - that would just mean extra wasted effort to screen the individuals, which "only" costs time and money.
The larger problem is that in order to do any real good you need an unbelievably low false negative rate. Let's take the 9/11 hijackers as an example: they were only about 0.00000667% of the population. Unless you could capture all but 2 or 3 them, you're still vulnerable to the plot unless you can get one of the ones you captured to spill the beans - at best you've just mitigated the plot. How realistic is it that ALL 20 (or 19 if you believe that Zacarias Moussaoui was not part of the conspiracy) of them could have been identified (let alone captured) using such a method, even given the expenditure of vast resources sifting through all of the false positives? Even if 4 or 5 of them manage to fall in the "negative" group or, alternatively, if they're able to slip through your second-level screening procedures, you still have a disaster on your hands.
It's not likely that you could get the accuracy high enough to stop very many plots by itself, I suspect.
I mean look:
A terrorist:
http://images.google.co.uk/images?hl=en&q=A%20terrorist&um=1&ie=UTF-8&sa=N&tab=wi
Not a terrorist:
http://images.google.co.uk/images?um=1&hl=en&safe=off&q=Not+a+terrorist&btnG=Search+Images
Problem solved. NEXT!
Deleted
I realize this is likely starting to sound old, but Cory Doctorow's Little Brother should be required reading for people doing something like this. His writings about the "Paradox of the False Positive" are enumerated there, but also in other sources:
http://www.guardian.co.uk/technology/2008/may/20/rare.events
(emphasis mine)
And, as others have pointed out, this system is likely to have a false positive rate higher than 1%.
At $8/hr TSA-guy isn't paid to think.
Engineering is the art of compromise.
Having personally used Multiple Data Mining techniques for several years now - It's not that Data Mining doesn't work, rather it's how its used. Data Mining is great at trend forecasting and if you're really good at what you're doing in it you can factor in probabilities of certain future events. The one key factor in data mining is a "Training Set" of Data to teach the machine(s) how to recognize the patterns. Since I suspect Terrorist come from every walk of life, every know nationality, and are using 1 off events this is throwing them a few headaches. The real key is to of course define what is normal, but if the rest of the world is as normal as are we here in the US they don't have a chance to pin point the Target Data (in this case people).
I would also suspect that the Terrorist Motives might be a key factor, but it's like pulling teeth to get any US Administration to admit that their foreign policy is screw up beyond belief, let alone something like a cruddy foreign policy might just result in cruddy foreign relations or popular uprisings around the world. If they did, then we wouldn't need data mining in the first place.
"May You - Live Long and Prosper in Interesting Times" -- by deweycheetham
Yes, because we need an aircraft that can fly faster than mach 3 above 80,000 feet to penetrate terrorist airspace and evade terrorist fighter jets.
They use Predator drones now because, well, they were not manufactured in the 60s, they cost magnitudes less for the same mission, they can perform more than one type of mission, they are unmanned, etc. etc. etc. etc.
The SR-71 was cool and all, but way too antiquated to keep around. Clinton made a good move killing the SR-71.
By the way, how did the SR-71 help anyway? It gave us zero insight into the Soviets...in fact, it did far worse by giving us false insight. It was a waste of money from conception to retirement. But...damn, it was cool...
IANYL, IANAL, TINLA, IANAMD, IANAP,
No. False positives "only" cost the government time and money. For the individuals falsely suspected, it could cost them their career, their relationships, their home, and their freedom, depending on how much "time and money" the government spends on them before realizing they are innocent. (If they ever do, since -- as shocking as it sounds -- there have been a few cases reported where individuals were detained indefinitely without charges, or even evidence.)
I am the man with no sig!
I said it before and I'll say it again: Any model that is built on 10 or 20 positive examples from a population of 6,000,000,000 is going to suffer from overfitting. Not just a little overfitting... I mean it's going to overfit like a mo-fo. There's just no way, and I mean NO way, to create a statistically significant test based on the data we have on who is and who is not an ACTUAL terrorist.
Music speeds up when you yawn, but does not change pitch.
I was wondering whether techniques of commercial data mining could be applied to environmental problems like emerging disease surveillance.
Well, of course they can. The question is how far is it from practical? I think, pretty far from being as practical as it is in business.
First of all, businesses have a great deal of object model in common: they have common concepts like customers, products, sales, brands etc., which form a common framework in which they can do all kinds of creative thinking, or if not thinking you can even discover relationships using some kind of machine learning.
Secondly, when you are dealing with business data, the most important events tend to be common events. The most important common event is when a customer buys something. When you talk about something like a new disease emerging, or somebody committing a crime like hijacking or bombing, the most important events are exceeding rare, but catastrophic. Therefore the connection between events we do have in abundance and the events we are interested in is tenuous, poorly statistically attested to, and in many cases pure conjecture.
Finally, a lot of what businesses use data mining for is tweaking marginal costs and revenue by shifting dollars that were already going to be spent from one place to another. Offer product A to this web visitor instead of B. Stock more of item X in the store rather than Y. If you really don't know a priori whether X or Y will sell more profitably, you probably aren't going to go too wrong.
In something like environmental monitoring, you create expenses that weren't already there. No, you can't drain this lake because the model predicts a 5% marginal increase in the probability of human cases of hantavirus in the area. To somebody counting on the economic value of draining that lake, that's a brand new cost that wasn't there before.
Same goes, even more so, to deciding somebody is a danger to society.
Now let me say that I have no doubt that data mining will lead to more terrorist being thwarted or captured, compared to doing nothing else. Of course so would a lottery, but I suspect that data mining is a great deal better at identifying good suspects than a lottery. However, it is for reasons I noted above not going to be particularly accurate, certainly not compared to probable cause. Furthermore, the marginal cost of false positives gained seems likely to exceed the marginal value of false negatives lost, if such things could be quantified.
Post may contain irony: discontinue use if experiencing mood swings, nausea or elevated blood pressure.
Why do people still stubbornly insist that flagging 'terrorists' was ever the reason for all of this data-mining? Don't people understand the hidden agenda is to develop detailed dossiers on every single ordinary US national?
'He who has to break a thing to find out what it is, has left the path of wisdom.' -- Gandalf to Saruman
Just compile a list of all the extremely illegal and unethical things you're doing as a government and find the groups of people most impacted.
Let those people simmer for 5-10 years under your asshattery and let cool. Presto! A tasty terrorist.. Bon Appetite!
The road between democracy and tyranny is paved with secrecy in the name of security.
It turns out that terrorism in western countries is a very rare thing, outside of a few hot areas like Spain's Basque area. This is very good, by the way.
Mining for rare event is extremely difficult. Bayes' s rule indicates that if in a database there are 0.01% actually suspicious events and your mining algorithms are extremely effective at 99% accuracy, then you still have an approximately 100:1 false positive ratio, which makes the mining still useless.