Software That Can Censor 'Sexual Images.' Or Not.

Halster writes: "Here's an interesting story on Newswire about censorware that detects excessive skintones in images, and implements blocking accordingly. What next?" What's next is (you'd hope) the realization that image-analysis heuristics are inherently limited, and not the best thing on which to pre-emptively base system-admin decisions. ( michael : That story is a company press release. For a much better evaluation of how this software works, see this Wired expose detailing the fraudulent nature of image-filtering "artificial intelligence," or an older review from Businessweek on Eyeguard.)

Comments

I'm on a Porn Collection Task Force

[David Wong]

"...Not only does eyeguard alert the network administrator, but it also disables the computer and takes a snapshot of the suspect image.."

My boss has installed this software, and is now forcing the entire office to surf for porn. These "snapshots" are sent directly to his hard drive, which is saving him the time of having to sift through thousands of non-porn pictures to get the ones he wants. Thanks to this software and the snapshot feature, my boss is able to accumulate pornographic images at 10X his previous efficiency.

Eye-T, Mr. Wilkerson thanks you.

hmm

[the_other_one]

What I want to see is an image filter that will filter out the clothing.

My favorite quote...

[A Big Gnu Thrush]

From the Wired article: "How do you tell the difference between a woman in a bikini in a sailboat which is not racy and a naked woman in a sailboat?" Touretzky asks. "The only difference is a couple of nipples and a patch of pubic hair. You're not going to be able to find that with a neural network."

Maybe our definition of obscenity is the problem.

Useful to alert admins, not to block

[Paul Johnson]

OK, its pretty obvious that obscenity is in the mind of the beholder, not the computer. So computers can't spot this stuff.

But I can imagine a program which tracks the average flesh tone score for pictures over time. If the moving average goes over a certain threshold then a dialog box pops up on the sysadmins screen telling him that Joe in cubicle 69 may be abusing company bandwidth, click here for a list of the suspicious URLs. Or, as it might be, sends an email to Junior's father. The key point is that this stuff can work as part of a monitoring system that uses human judgement for the final bit, rather than being a blocking solution.

Companies do have a legitimate need to monitor this stuff. Quite apart from the abuse of company resources, companies who allow employees to download and view sexually explicit materials can find themselves on the wrong end of a big discrimination lawsuit.

Paul.

The real dark side of this

[jabber]

Image recognition, refined enough to filter porn, will not be around for a VERY long time. I'm not that imaginitive, and I can easily picture all sorts of "unnatural" positions which an automated system would have a hard time recognizing as porn. :)

It will take an AI with the understanding of what "porn" means, with an appreciation for the human body's full range of motion, and with the comprehension of the latest fetishes - else National Geographic and CNN.com will find themselves filtered out of libraries and schools. After all, what is the difference between an image of a 'man riding a horse' and that of a 'man riding a horse'?

But the research being put into this sort of image recognition has an even seedier and more sinister side. It can/will filter based on LOGO. That's right.

Imagine Time-Warner/AOL being 'unable to connect' to sites which feature their competitor's logos.. Imagine ISPs who show Reebok ad banners suddenly disabling links to pages that display the Nike "swoosh". Imagine your favorite web-site suddenly not letting you click through to any other site that does not proudly wear a "VA" on it's 'sponsors' page.

And all this technology is being developed... (oh, say it with me) "In the name of the children!". BS - all the children I know would get a kick out of looking at porn, and are being damaged more by advertising than by sexual content.

Personally, I think we should assist in the development of this technology, and make sure that it only filters on Red Maple leaves on white backgrounds! Blame Canada!! Hooyah!

Of tanks and sunny days

[jabber]

This whole thing reminds me of an ancedote on image recognition using neural networks in the early 90's.

The military was developing a neural net to optically identify tanks hiding in brush and trees. They trained the net on photographs, and noticed exceptionally good performance. Then they tried a new set of photographs, and the success rate of the network suddenly fell flat.

It turned out that in the first set of photos, all the 'with tanks' pictures were taken on a sunny day, and all the 'no tank' pictures were taken on an overcast day. In the second photo set, this was reversed, and the network continued to correctly tell sunny and cloudy apart...

So I have to wonder what this new development holds in store. Will we be graced with filtering software which discriminates between gaudy, tacky motel room backgrounds rather than drunk co-eds on crack taking it in the ear? Will it be intelligent enough to tell if someone trims their hair, and is therefore a professional, versus those fat and hairy amateurs?

Hey, maybe it will finally be able to tell if the subject of the questionable photo is in fact "barely legal", or has the stretchmarks of a few litters of puppies to suggest the contrary.

Your point on skin color is well taken, and opens up the question: What does ALL porn have in common that can possibly be quantified and filtered? The answer, I'm afraid, is nothing at all.. One man's porn is another mans art.

Aside: I, for one, didn't consider the Maplethorpe photos to be obscene. They weren't artistic either. They were just 'for shock', to cause a furor, and get more attention through objection than through inherent value.

My strong suspicion is that whoever is behind this effort is a lot smarter than they seem at first glance. They are trying to bleed the "Religious Right" of money by getting 'upstanding' Bible-thumpers to fund this research - all the while knowing that it can not possibly be successful.

"Oh look, that girl is wearing a schoolgirl outfit - this is a porn pic, filter it!" -- so much for all the Brittney Spears fan pages..

Best filtration software...

[ocelotbob]

I have the best solution to the Censorware problem, plus it'll make a lot of people very happy. First, in the rich, upper crust neighboroods, you advertise for realtime filtration of Bad Stuff to protect the whelplings. Next, you advertise in and around colleges such things as "Make Money Viewing Porn". You pay these students about $6.00/hour. Now you put all these students in front of computer terminals, and hook them up to heart monitors. Any time someone subscribing to the service wants to view a page, it's first shown to one of the random college students. Now, if their heart rate rises once they see the page, you know that the page should be filtered.

Nothing new, just a new scam

[Old Man Kensey]

I remember speculation about filtering web content by checking the skin-tone levels of images as far back as 1996. At the time everybody more or less decided it was too impractical and gave up on the idea. The standard points about non-flesh fleshtoned objects and large amounts of non-pornographic flesh were made then too.

This is a particularly disgusting (to me at least) instance of the "for the children!" canard. Now instead of politicians using it to achieve their aims, which is bad enough, we've got a company using it to bilk panicked consumers out of their money.

And of course, just as with the quality of our politicians, we Americans have only ourselves to thank for this. If people weren't so damn gullible, companies like this would never sell a dime of product (of course in this case it's questionable whether what they have constututes a "product", but the point stands...)

What's needed is people willing to stand up and say "Yes, damn it, I do support porn on the Internet, and the easy availability of information on things like bomb-making and lock-picking, and if you don't like the speech I support, TOUGH SHIT. You don't get to pick and choose. If you want free speech, you got it. If you don't want it, go start your own damn country and LEAVE MINE ALONE."

But what are the odds of that happening?

Heuristic analysis

[jabber]

All forms of naked women are to be filtered, except when their arms are missing, in which case it's Venus de Milo, and therefore a bona fide work of art.

Clears the way for amputee fetishes, I think. :)

Boticcelli's Venus, the image of a naked woman coming out of the surf, that has been used as the box art for Adobe Illustrator (IIRC) would of course be flagged. She has nipples and a 'patch of hair', as do most nudes painted during that time period....

Hell, the Sistene Chapel ceiling is offensive, it shows Adam (naked youth) and God (Old man) touching fingers.... There's a bunch of naked little boy cherubs flouncing around them to boot. What horrific kinkiness!!

Only for white-folk?

[Anonymous Coward]

What about us black folk?

No, I am not trolling. This is seriously flawed. Not to mention stupid.

Detecting Skin tones

[DeadSea]

As part of a project for a multimedia class that I took as a Senior in college, we had to write software to count the number of people in an mpeg video. It was a very open ended project and we weren't expected to be able to get the right answer all the time.

The professor suggested that we start with skin tones. He pointed us to research that tried to pick out the parts of the spectrum considered "skin tone". There were some simple algorithms that were suggested. We did this and it worked decently well, but there were a lot of things that looked like skin to it. Especially light colored woodwork.

An algorithm like this may be able to filter a lot of stuff off the web. But it will filter a lot of other stuff too. I can also think of 100 ways to fool it. The easiest being put images through a color filter before posting them, or post them in black and white. Other people have pointed out that it will filter portraits and other shots of humans that arn't porn.