▲ 72 points
back
79 comments
Gizmodo rewrote http://www.telegraph.co.uk/technology/2017/12/18/artificial-... and didn't include some interesting details.
Was it Gizmodo that posted the doctored version of James Damores memo where all references were removed? Clearly not a reputable place to get information from.
Color me flabbergasted.
The UK police are dealing with a scandal right now in that they tend to suppress electronic evidence proving a suspect is innocent and try to prosecute anyway. The last thing we should give them is another electronic capability.
http://www.telegraph.co.uk/news/2017/12/19/met-orders-review...
So, the Met wants to train image recognition AI to find child abuse images? I'd be surprised if there aren't some interesting ethical questions around this.
To begin with, the way image recognition algorithms work -they learn from examples- to train an image recognition AI to identify images of child abuse you'd have to provide it with a substantial database of such images. I'm pretty sure that's illegal, although of course the Met probably has some sort of exception for the purpose of fighting crime. But, they still have to keep around a big database of child abuse images - and update it regularly, to do this job. It's kind of... icky.
Then again, what about victim protection? Besides the fact that keeping such a database is risky because it can always fall into the wrong hands (it only takes one misconfigured database server) it's also a characteristic of the most advanced image recognition algorithms (conv-nets) that their models can be used to generate new images of the kind they've learned to identify. So presumably, if the model itself fell into the wrong hands, someone could start generating new child-abuse images from it (low-fidelity and not really very useful, I imagine, but still).
So I wonder if the Met has addressed such ethical issues, anywhere.
The only thing new is that the child pornography pictures are being fed to an ML model. The police have been keeping and distribute photos of icky things for as long as pictures of murdered people has been useful as courtroom evidence, and Interpol and the FBI have kept a database of child pornography since the early days of microcomputers to track down rings of child pornographers.
As far as generating new kiddie porn, before there were ML models, there was Photoshop, and that first released in the 90's. That there are more advanced methods of generating child pornography is not interesting or new under the law (depending on your jurisdiction).
"Fun" fact: Facebook/Instagram/Snapchat have giant repositories of (tagged!) child pornography however before we get our pitchforks out, it's used to keep child pornography out of the system.
Unfortunately with current technology there's not really a realistic way around having the actual images; hashing only goes so far, and just like (performant) homomorphic encryption, if you were able to come up with a way to do this without actually needing the pictures, the world will beat a path to your door.
Who wants to be the developer to train a neural network on child porn? You couldn't pay me all the money in the world twice over to take that job.
> But, they still have to keep around a big database of child abuse images - and update it regularly, to do this job.
IIRC, they already do this, for tracking the distribution and source of such images.
> It's kind of... icky.
No disagreement there.
> their models can be used to generate new images of the kind
Tools like Photoshop can already do this, and likely to a higher quality than running a conv. net in reverse (so to speak), so I doubt that's a big concern.
Law enforcement already has tons of child porn lying around used primarily to frame people. Of course similar to drug planting it is extremely hard to defend because no one wants to risk defending someone charged with that. There are exceptions and I'm not saying people like Jared from Subway aren't guilty. There are many many many lesser known individuals that are hushed up using child porn.
So I'm basically saying there is no ethical problem (from their perspective) because feds/cops already deal with this all the time and plant this on people like drugs.
Not unexpected. It's surprisingly difficult to tell your arse from your elbow:
Anything can be deceiving. https://imgur.com/a/V6GBV
But... I... I just want to know the answers. That reset quiz button... It’s... Why?
lol! The missing "submit" and only a "reset" button also drove me nuts!!! That page actually does some javascript code, which is not currently working. (you can see this when you click on one of the options while in console/devtools).
Its a ruse, they just want you to look at random people's backside and think that it's a elbow. The modern day version of scanning your ass and printing it lol
From the source: false, true, true, false, true, true, false, false, true, true, true, false, true, false (go left to right, top to bottom)
Yeah the human brain i capable of finding sexual imagery in the weirdest of places.
As such i am unsure if this "AI" is failing or has become all too human.
Elbow grease has taken on a whole new meaning to me
A bigger question is why does the police spend time looking for pornography? What type of society is the UK developing into?
This is from 20 years ago, the idea was to scan a selection of laptops going through border control on Eurostar, although it only worked on PCs:
We (the UK) recently experienced a scandal where a celebrity was allegedly a prolific paedophile. After his death it transpired that numerous incidents had been reported and covered up by the police and the BBC (his main employer).
So now, guess what, the same police and BBC have transformed to be holier than thou organisations free to cast judgement on anyone they don't like because this will absolve them of any guilt regarding past failures (obviously).
Isn't this specifically intended to identify images of child abuse though, rather than rifle through the average users pornhub cache?
Not a great one, but that's not new. We've been getting incrementally more draconian surveillance and laws for some time.
> The police force already leans on AI to help flag incriminating content on seized electronic devices,
This is people who're being investigated for possessing images of child sexual abuse. This might be because they've accessed a website known to distribute such images; or they've created imags of a child which were found.
Police need to search all the devices that person owns in order to find all the images. They need to do this to protect children who are still being abused; to detect unknown abusers; to add images to the various IWF etc lists; and to ensure the correct prosecution and sentencing approach is taken.
No, this was for after work.
So they can bully those who they don't like the disgraced bob quick's (senior counterterrorist cop who had to resign for security breaches) vendetta against a tory MP is one case.
> With the help of Silicon Valley providers, AI could be trained to detect abusive images "within two-three years", Mr Stokes said.
If the goal is to get something so high quality that the police won't ever have to look at it, I doubt it. You have Google and Facebook employing people to deal with user reports about this stuff, as well as ad companies manually checking websites.
There would be a significant benefit to just being able to reliably get rough numbers plus some examples of the worst pictures for an expert to grade rather than having the task of searching 10000 horrific pictures to see if any fall into certain categories.
I doubt that's the goal (or at least if it is, it is shortsighted). There's still a benefit (to the police) of being able to say "Here's 100,000 images, of which we strongly suspect some percentage are abusive. Show me those images which are _probably_ illegal", even if just in time spent.
I don't get it. Who thinks it's a good idea to fully automate anything like this? Regardless of how accurate the algorithm is a human should look at this before any action is taken. As long as the aren't false negative this is probably a huge win. Even if we have 5% false positives that hugely cuts down at what people need to look at.
So desert pictures looking like child porn doesn't matter much. Child porn looking like something else is much worse.
Edit: awful typo "lol" -> "look"
well, they can kiss my shiny metal: https://www.flickr.com/photos/wiredfool/901581193/
And then there's the classic Weston photograph: https://en.m.wikipedia.org/wiki/Pepper_No._30
(Edit, IOW, not a new problem, photographers have been pushing at this sort of issue for a long time)
I suspect part of the problem with this approach relates to data set construction. Even though models these days should theoretically be able to handle this task, there are clear ethical concerns and practical issues with making datasets of illegal imagery large enough for training. It really raises the question - is creating a dataset like that ethical, assuming the intentions are to stop further abuse and dissemination?
There might be work arounds (like training one model for nudity and another for age) but such approaches are almost certain to have "suboptimal" performance as compared to a single model trained on a relevant dataset. Maybe something like that is the cause for the performance issues discussed in the article.
Honestly, if they stored evidence collected from previous cases, the data probably already exists in some form. As long as the data was closely guarded internally, I'm not sure if there would be an ethics problem using this for training.
The biggest problem with obscenity detection, though, is getting the context right. The AI might be able to get to the point where it can detect "naked human" at a good percentage level. At the moment, however, I doubt it could detect whether the naked human was considered obscene in current culture, eg: the difference between "child pornography" and "famous Vietnam War photo" as alluded to in the Gizmodo article. So no matter how good their model gets, without further refinements in AI it would only be good for a "first pass", I would think.
Ethical is a tricky issue here. Not everyone agrees that mere possession of such pictures is unethical.
This still leaves the obvious legal problems of having a database of illegal pictures, but we already let law enforcement do otherwise illegal things (including use of illegal pictures in some contexts)
Having grown up in Saudi, I had heard that they had (literally) the same issue when MMS was originally introduced many years ago - but they have a lot more desert and are simply looking for porn, not even anything abusive.
Given that children are forced into marriage in Saudi, it would seem strange for them to care about images of the abuse they condone. A small bit of consistency at least.
> Stokes told The Telegraph that the department is working with “Silicon Valley providers” to help train the AI to successfully scan for images of child abuse
AI requires large image sets to do the training. These images are illegal to possess. I wonder how such training gets done without violating the law. However, I do not wonder enough to do web searches about this matter. I am probably on enough government lists already
The Silicon Valley television show had the best parody: a deep learning program trained to find hot dogs for a foodie program repurposed for a phallus porn detection program.
There are some training techniques that rely on negative examples to improve classification, though I'm not sure they exist for images
Well, this can somewhat be accomplished by classifying exactly what an image isn't, but as you can imagine, that is (in the general case) a combinatoric nightmare.
Anyone else have an Arrested Development flashback? :)
Close up, they always look like landscape...
That headline. FFS. It's either child abuse or pornography. It can't be both.
Sure it can. One person's abuse can be another person's pornography. Disheartening but no less true.
The Telegraph headline uses the correct language. http://www.telegraph.co.uk/technology/2017/12/18/artificial-...
perl -e 'print "yep, is porn\n" if $filename =~ /(Porn|Sex|XXX)/;'
Should spot the 99% of cases.
perl -e 'print "yep, is porn\n" if $filename =~ /(.avi|.mp4|.wmv)$/;