Computers that know what porn looks like

After I went a little overboard with the longest Geek Talk... er, Nerd Talk ever in the history of the Dirty Little Secrets podcast (episode #99), I thought it appropriate to do a little Nerd Talk that stays on the whole "porn" topic.

There have been many filter applications that have been developed over the years that try to prevent people from accessing (intensionally or by accident) pornographic websites. The filters work in a number of different ways. One of the most common ways is to look for keywords that are considered inappropriate. The keywords could show up in the URL, or within the content of the page.

This method is a bit of a problem because sites that do no contain any pornographic material but have "inappropriate keywords" are blocked. This site, for example would get blocked in a heartbeat. XXXchurch.com. The #1 Christian Porn Site.

Solution? There are a few - like having a whitelist of all the sites that might be flagged but are trusted sources - CNN.com, MSNBC.com, XXXChurch.com. But that list would go on forever.

It looks like the best alternative would be to get a computer to learn how to look at porn. In other words get the computer to look at each image before it is displayed to the user and determine if it is pornographic. Enter Pornographic Image Filter. Looks like the developers over at SafeSquid have cooked-up an algorithm that can tell if an image has naked people in it.

"Pornographic Image Filter can analyze an image in real-time, and identify the ones that are pornographic in nature. It analyzes the graphical content like skin tone, contour, etc. to identify a pornographic image. [...] Although it is about 85%-90% accurate, it acts as a good deterrent."

It looks the program runs on *unix machines for enterprise-level content filter servers, so it doesn't look like you'll be able to install this on your home computer. I'd like to see if this technology will soon be implemented at the consumer level.

Although, if you're using a Mac (recent convert here), you're running a unix system. Hack it! ;)

Brian Baute wrote on August 21st 08 at 06:01PM

That's awesome - can we can expect that feature in the next release of X3Watch Pro for Mac?

slash wrote on August 21st 08 at 10:11PM

FYI, I have been using SafeSquid 20 user free edition for quite sometime. This edition has all the features of the commercial edition, and can be installed on your home computer, provided it has Linux. Their website does say that they are in the process of releasing a Windows version soon.

The Keyword filter in SafeSquid, does not block web pages based on a single word like 'xxx' or 'sex', but only blocks it when it find multiple 'phrases'. It does not block articles, news, discussions, educational and medical websites.

Another advantage is that it allows you to use open source software like Clam Antivirus and URL Blacklist like http://www.shallalist.de/

There are many other useful features.

slash wrote on August 22nd 08 at 10:14PM

SafeSquid Personal for Windows (Beta) has just been released. See http://www.safesquid.com/html/viewtopic.php?t=2819

Ian wrote on August 23rd 08 at 01:30AM

If it runs on *unix, it will run on my linux machines. Awesome.

SammyK wrote on August 23rd 08 at 02:13AM

@Brian - I don't think that will be added to X3Watch any time soon. Sorry man.

@slash - Thanks for the link! That's so cool how the day after I said, "I'd like to see if this technology will soon be implemented at the consumer level," they release a Windows consumer version. Weird.

@Ian - What flavor of Linux do you run?

Ian wrote on August 23rd 08 at 08:22AM

@Sammyk - Ubuntu 8.04. Yes, i know *everyone* uses it, but thats because its a good distro. Also there is a fairtrade cola drink called Ubuntu. How awesome is that?

SammyK wrote on August 23rd 08 at 02:09PM

Nice. Ubuntu is definitely my favorite distro.

Suricou Raven wrote on September 25th 08 at 02:54PM

I would like to see some better evaluation data. If it's going on skin tone, there isn't really any difference between porn and the closeups in the family photo album.

Sean wrote on September 25th 08 at 09:55PM

The article says that it is about 85-90% accurate, so there are 15% possibilities of false positives. You can fine-tune its 'Threshold' sensitivity, see - http://www.safesquid.com/html/portal.php?page=80