YouTube Video-Fingerprinting Due in September

Tech.Luver writes "The Register is reporting on Google's statement to a presiding judge that video-fingerprinting of YouTube material will be ready in September. The development is required to head off a three-headed suit against the company, currently being debated in a New York City courthouse. The system will, according to Google, 'be as sophisticated as fingerprinting technology used by the Federal Bureau of Investigation.' From the article: 'As Google told El Reg in an earlier conversation, the company already has two systems in place for policing infringing content - but neither are ideal. One system allows copyright holders to notify Google when they spot their videos on the company's sites. When notified, the company removes the offending videos, in compliance with the American Digital Millennium Copyright Act. A second system uses "hash" technology to automatically block repeated uploads of infringing material.'"

Comments

Hard AI ftw

[UbuntuDupe]

Earlier I had joked about Google's claim to be nearing a system that lets them check for copyrighted works. I said that they're basically claiming to have solved a hard AI problem.

Others pointed out that, no, it's not a hard AI problem to just compare some kind of checksum of the video against a set of banned checksums. That's true. But what about once people know they're using this system? They can just trivially re-encode. Perhaps add a scene break here or there, and totally mess up the fingerprint. To prevent that, it seems, you would need to solve a hard-AI problem: that is, be able to determine if an arbitrarily-encoded video appears to a human to match some copyrighted work. It would have to be robust against minor scene shortenings and lengthenings, scene breakups, color gradients laid over the video, etc.

Anyone know how difficult this program is to circumvent? (Just hypothetically -- not advocating criminal activity here.)

Re:Hard AI ftw

[sholden]

So I guess it doesn't matter if it then blocks a video which has taken a couple of seconds of video from a TV show in a "Review of Episode X" video post, and just happens to grab one of those keyframes?

Of course a little bit of coding and you have a program that takes that 10 minute video, splits it into 10 1 minute videos and uploads them. The ones that got rejected it splits into 10 6 seconds videos and uploads them. Rinse and repeat until you have however small an set of rejections you asked it for. Then it cuts out just the necessary fragments of videos (replacing them with the last good frame or something?).

Of course that can be worked at google's end by adding a delay to the report rejection step, and by banning those who get lots of rejects.

Re:Hard AI ftw

[DDLKermit007]

Sorry, but fingerprinting the video is damn near impossible. Closest any image recogniton app can go is similar scene right now. Some can do logos, but were talking low resolution videos. Pretty much they'll do a file checksum, and that'll kill most reuploads. For the determined. All that'll be needed is a simple re-encode. That'll completely change any checksums involved. They won't seriously spend that much machine time on checking if they are allowed to make money off a video. They'll do just enough to say they did something, but not so much you can't find a way around it.