New Photo Fraud Detection Software

An anonymous reader writes "CNet is reporting that Hany Farid, Professor of Computer Science and applied mathematics at Darthmouth College, has developed a new version of his Image Science Group's photo fraud software now in use by the FBI and large media organizations. The current software is written in Matlab, but the new version will be written in Java making it much more readily available to local police and smaller media organizations. From the article: 'I hope to have a beta out in the next six months,' Farid said. 'Right now, you need someone who is reasonably well-trained to use it.'"

Comments

Same shot from a different angle...

[Tsar]

Note that the submitter identifies the product as "photo fraud software" not "photo fraud detection software." This is quite apt, since the application will obviously cut both ways. Someone cooking a photo could simply run each version through this software, making minor tweaks until their "improvements" pass its inspection. If the software works the way it appears to, it would be the image manipulator's equivalent of a spell checker.

Re:open source?

[NathanBFH]

Negative. The GPL doesn't have such a clause, but it is easy enough to modify the GPL (or any other license) to allow distribution of the source with the binary, and then restrict distribution of both. Granted, you may be using a different definition of the word 'open' than the software owner is. This scheme is more open than a closed-source solution where I sell you a binary with rights only to use the software and absolutely no rights to the source. While it may offend your OSS senses, it is, I think, a valid use of the term open source.

If you really want to take the term 'open source' to the extreme, I could argue that even the GPL fails to meet some level of openness. The GPL restricts use of its source code on several accounts.

Fallibility

[Mikey-San]

From the article:

"Right now, you need someone who is reasonably well-trained to use it."

I would like to hope that if this software is going to be used for anything of consequence, that someone reasonably well-trained will always be using it. A system is only as good as its operator, ultimately.

It's a fraud, because my blackbox software says so

[Jivha]

As other commenters have already pointed out, the confusion over its open-source nature(is it or is it not?) is critical. Without the source code/algorithm being open-source and freely accessible to the public how can one trust its "judgement"? In a legal situation, an accused will always question the accuracy of the algorithm and the software.

On a different angle, I wonder how soon before such detection capabilities will be available to consumers either as an installable plugin or web-based feature. Imagine being able to verify the authenticity of any picture on the web, ranging from that nude shot of your ex- to that impossibly perfect low-light picture taken by your photography class buddy ;-)

Re:It's a fraud, because my blackbox software says

[jimicus]

What you're saying is a rehash of the argument that "Information security isn't secure when the algorithm's security depends on its secrecy".

Thing is, does the same thing hold true when you're talking about detecting fakes (say), as opposed to building strong encryption? If I announce "Well, we can tell this photo isn't genuine because this part which shouldn't be in focus is", I've effectively announced to any potential fraudsters who might be listening "OK, folks, you need to learn to get your focusing correct".

Realistically, the only way such an algorithm remains secure is if it cannot be beaten even with a full understanding of how it works - and I would ask if such an algorithm even exists yet. If the algorithm is anything less than 100% effective, chances are it doesn't.

Re:A matter of time

[pyite]

It uses the same algorithms in a slightly different way: instead of checking for the signs of forgery it finds the tell-tale signs of modification and then reverse-modifies them to "what-should-be-there" to make an "original" modified image.

What makes you think this is possible? Let's say I have a set of 15 numbers, {1, 2, ..., 15}. Now suppose I consider each of those numbers modulo 3. I now have a much smaller set of unique numbers, but multiple numbers in the original set map to the same new number. I can't solve x (mod 3) = 2 uniquely for x. A whole bunch of numbers satisfy that equation. Taking this back the fraud detection standpoint, it's likely that while you might be able to algorithmically generate a fraud, it won't look anything legitimate to the human eye. Much as the collisions found in MD5 are not meaningful because there is no change of mistaking a real message with a fake one with the same hash. So, your thought about reverse modifying is probably unlikely. That's not to say that other weaknesses in the algorithms won't be found.

Re:Government Funded

[Anonymous Coward]

Who said anything about the GPL? Professor Farid certainly didn't.