Despite the headlines, Rudd's online terror takedown tool is only part of the solution The UK government launched its swish new tool to fight online extremist content to much fanfare this week – leading news bulletins and generating reams of coverage. But it also faced a whole host of criticism and concern. Underneath the claims made by Home Secretary Amber Rudd yesterday, there are many questions yet to be …
"He said that the biz had balanced false positives with performance to tune the algorithm to be able to detect 94 per cent of Daesh propaganda with a 99.995 per cent accuracy."
Did you ask him what he meant by this? AFAICS if it correctly detects 94% of videos it's 94% accurate. What, then, if anything, is this 99.995% figure? Unless he has a sensible explanation for this then I wouldn't trust any figure he provides. It's simply marketing gobbledgook and deserves to be treated an the same way as any other garbage spewed by marketing mouthpieces.
^ that one. Though given the sets outlined in the article I'm not sure they're actually assessing accuracy, it sounds more like the 99.995% figure is specificity. To get accuracy they have to factor in how common positives and negatives are.
For the original question (and my earlier reply went astray, so this is the even shorter version): there are different figures of interest. Changing one part of your detection may change some more than others, and overall success rate is still dependent on how frequent positives (in this case, terrorist videos) are relative to negatives (cat videos).
Sensitivity. How frequently the test reports positive given only positives. 0.94something here. You can trivially make it 1.0 by always reporting a positive.
Specificity. How frequently the test reports negatives given only negatives. 0.9995 here. Trivially 0.0 by always reporting negatives.
Obviously in most cases you can make a tradeoff of one for the other. Receiver operating characteristic curves are usually used to illustrate that. Flip a coin every time and both your specificity and sensitivity are 0.5 (as is accuracy, which is usually more complex), any random choice that doesn't look at the input sits on a diagonal line on the ROC plot.
Accuracy is a composite, but a more complex number. As Mycho said, Accuracy = (True Positive + True Negative) / (Positive + Negative). Positive plus negative is just all your samples, but true positive depends on how many positives there are and sensitivity, true negative depends on how many negatives there are and specificity, so your accuracy figure depends on the ratio of real positive and negatives in the input. If the vast majority are negative, then sensitivity has very little effect on accuracy, if most are positive then specificity has very little effect.
This tends to come up in medical screening, https://en.wikipedia.org/wiki/Sensitivity_and_specificity two numbers not much talked about are positive predictive and negative predictive value. Both of these mean if you select samples (videos, people) from the general population (youtube, a screening population) based on a test, how much does the ratio of positives (terror videos, people with a particular disease) change relative to the population. A rare disease and a test that is not sufficiently specific may still end up selecting a majority of people who do not have that disease, even if sensitivity is incredible.
Too late to edit:
Specificity. How frequently the test reports negatives given only negatives. 0.9995 here. Trivially 0.0 by always reporting negatives.
For specificity should be: "Trivially 1.0 by always reporting negatives". On the flip side, always reporting negative regardless of input gives 0.0 sensitivity, while always reporting positives gives 0.0 specificity.
I remember that the Blair / Brown government was just as likely to trot out the "But Terrorism" defence every time that introduced yet another piece of intrusive legalisation.
At this point I am forced to feel that this attitude is endemic in the whole UK government machine and the front bench of the day only puts a face to the mess.
At this point I am forced to feel that this attitude is endemic in the whole UK government machine and the front bench of the day only puts a face to the mess.
I am well passed that point.
We've not had, nor are likely to, a Liberal party government to test the rule on, and unlikely to be able to test the 'new old labour' current alignment before it gets replaced by the new new labour corporate trojans.
The moment they start using this I'm going to get a daesh flag and upload pictures of my cat playing around on it.
Lets see how accurate it is then.
Bad stupid form of backdoor censorship that will be abused.
Edit: On second thoughts I'll just superimpose all politicians speeches onto a daesh flag and report them, eventually this AI will consider all politicians medieval terror bastards and block them from the web.
I assume Google will manually review any flagged videos either for reporting to the relevant authorities or to flag them as not terror related.
If somebody presses the "nah it's another blood cat wearing a tea-towel" button, surely that video will go in to training data further improving the algorithm.
This is only a problem if Google don't manually review or if further training doesn't improve performance.
This is only a problem if Google don't manually review
And the people doing the manual review read Arabic and are experts on middle eastern politics?
Or is "review" just flag anything with curly writing in it?
Suppose you were a reviewer for Egyptian Google, could you tell the difference between a Britain First rally and a Remberance Day parade or a Scout troop? Could our algorithm?
Facebook's child porn algorithm blocked the picture of the Vietnamese girl whose clothes had been burned off by Napalm and their manual review confirmed the decision.
Or perhaps it wasn't their child porn algorithm, perhaps it was their corporate media profile algorithm responding to their SLA with Dow Chemical ?
No doubt, there won't be one. At least, not one that matters. My guess is it might look a bit like this:
"To appeal, send a letter to this obscure government department (email not accepted). Within 30 days we will reply whether to allow the appeal. You then have six weeks to submit any further information for the appeal, and we then will have a target to decide 25% of all appeals within six months of the original takedown date (up from 18% last year). Please note all appeals are under an NDA and it's a criminal offence to reveal whether you have appealed or what stage the appeal is currently at, even when we've decided in your favour (which, let's face it, we probably won't)."
Dare I presume that uploaded videos must not include acts of beheading, pleas to 'come and work for ISIL, the friendly international terror organisation' or other Daesh propaganda .. and doesn't that rule out most Daily Mail uploads as well then?
On the presumption that most video content from Daesh is going to be rich in violence, rhetoric and the promise of paradise surely the ASI logarithm has a pretty clear criteria to start with, and therefore the 99.99etc% reliability claims are indeed sound?
1. The machine learning algorithm gets trained using an existing set of known styles of video.
2. Deploy to live and start rejecting new videos which match criteria (I assume maybe the accounts also get flagged as potential producers of dodgy material and get suspended, rather than having special forces sent to where the user is logged on?).
3. Dodgy uploader uploads their latest video using fresh account.
4. Dodgy uploader gets told their video has been rejected.
5. Dodgy uploader contacts their film director/editor to tell them their latest style has been compromised.
6. Dodgy director/editor moves on to a new style, and gives video to uploader. Go to 3
7. When a human sees an example of new style, reports it. Goto 1
8. So eventually, based on the above rules, we will have every publishing/film style ever used by these people? Black & White silent movies, French New Wave, Manga, etc. etc. and then they will have to give up? Is that the theory? Looking forward to their version of Carry on up the Khyber btw.
That will be the initial training and trial. Working on historical data. Once a few passes with refinements added manages a very low hit rate from undetected videos then the fun starts,
Full scale whack-a-mole.
This just raises the bar so that a 2 minute amateur video from a cellphone now gets screened out and more effort is needed to publish.
It was this rabble's predecessors and their mates abroad who started the whole thing in the first place. Listen to We Didn't Start The Fire by Billy Joel for some examples of what a foreign policy which has a Geordie Saturday Night attitude to someone spilling their allies' pint does to the world.
Yet, somehow, it's the ordinary people who must sacrifice any illusion of privacy and free speech while having their genitals inspected at airports by an ape with a three digit salary and an IQ that begins with a decimal point to clean up the mess they made and are probably going to continue making with impunity. Bastards.
The "how they're going to tackle this" isn't quite as important as the "why they now have to" which should, all things being equal, add a little perspective to the debate. Of course, it won't.
I am guessing the creators of this AI have not checked out some of the more nefarious black hat forums on the internet where there are entire threads on how to bypass the copyright filters used by Youtube and the like to upload copyrighted content without it getting flagged.
All that will happen now is that ISIS will start to use those black hat techniques to fool the AI and by pass the checks, resulting in ISIS videos making in through while lots of false positives will be held for moderation.
We wouldn't have ISIS if Saudi Arabia hadn't lost control of Al Queda over a disagreement over why the targets should be chosen by America, rather than be America.
And we wouldn't have that if it wasn't for Iran having a leadership Saudi Arabia don't like.
And we wouldn't allow that unless they also had a leadership we didn't like.
We they wouldn't have if we hadn't deposed the leader before last and imposed our own mad dictator
And we wouldn't have that popular leader if we hadn't run the country as our personal oil field
.....
And we wouldn't have that if Ug the caveman hadn't worshipped the moon in the wrong way according to Blurg the caveman.......
Ah yes, the age-old "my god has a bigger dick than your god" issue. For all the rhetoric about separation of church and state, the disorganised religion mob do still tend to sway the decision-making process quite a bit while, it must be pointed out, paying none of the tax to actually implement those decisions.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
