YouTube Filtering Is On-Line

ghostcorps writes "After months of promises to IP-holders, the long-awaited filters system for YouTube has gone online. The new system will make it easier, the company claims, for copyrighted clips to be removed. 'YouTube now needs the cooperation of copyright owners for its filtering system to work, because the technology requires copyright holders to provide copies of the video they want to protect so YouTube can compare those digital files to material being uploaded to its website. This means that movie and TV studios will have to provide decades of copyright material if they don't want it to appear on YouTube, or spend even more time scanning the site for violations.'"

It's A Shame They Won't Take the Offer

[eldavojohn]

What I found most interesting comes from the beta announcment [youtube.com]:

Copyright holders can choose what they want done with their videos: whether to block, promote, or even--if a copyright holder chooses to partner with us--create revenue from them, with minimal friction. YouTube Video ID will help carry out that choice.

Because I'm certain Google realizes that a lot of these copyright holders are sittin' on a freaking gold mine here.

I guess that's the sad thing though, it's no longer the people that made this stuff that own the copyrights. It's huge corporations. This goes for sound and video. Do you think any of the big studios care about artist exposure? They don't care about building a fan base, they care about profit margins.

I personally would like to see Google help users approach and push the limits of fair use of sound and video. I think that a lot of artists would be open to their work being displayed in a tasteful manner without the full work being put online. I also think that the usually low quality of YouTube is a good reason to allow this and that if copyright material is found, they should investigate either shortening it or degrading the quality so that viewers get a taste. What's more, putting a link to sales of the item would be basically free advertising.

I feel especially sorry for the people who build movie montages with unpopular songs [youtube.com] for I have watched many of them and purchased a DVD & CD from seeing the two. After watching that particular video, I rediscovered the genius of Sergio Leone after a fan posted that video with one of my favorite bands, The Arcade Fire. Sure, it's just anecdotal evidence but I still view that as original art & innovative.

It's truly a shame that copyright holders are throwing away what could be a beautiful & profitable relationship with fans.

How easy is circumvention?

[AmIAnAi]

Presumably they are creating fingerprints from the original material and comparing those against uploads. It would be interesting to know how well this copes with different codecs and frame rate changes.

Or do they wait for the uploads to be flagged as infringing and then do a dumb binary compare to prevent deleted files being uploaded again.

Circumvention Ideas

[CheeseburgerBrown]

1. A filter that shifts 70% of pixels one pixel to the left.

2. A filter that munges the rows of pixels around the frame area, distorting the video fingerprint without affecting viewing quality.

3. A filter that randomly inserts the Goatse man for a Fight Club-like single frame.

4. A utility that uploads the clip backwards, and then a browser-player that automatically time-remaps it forward for playback.

5. A watermarking process designed to distort the video fingerprint while remaining invisible to non-AI viewers.

Okay now -- code it.

All material

[Nosklo]

because the technology requires copyright holders to provide copies of the video they want to protect

Wait. That means google will pretty soon have almost ALL COPYRIGHTED MEDIA in its servers?
I, for one, welcome our new media-holding overlords.
There's a lot of money to be made with this material, besides searching youtube. Even without releasing it.

Re:How easy is circumvention?

[imgod2u]

Wavelet approximation does a pretty good job of being independent of framerate. Codecs just need to be decoded into raw information, and then analyzed. Hell, a simple FFT of the video, normalized to a certain framerate, would also do a bang-up job of filtering out 99% of the videos that don't match. The staggering amount of processing power required for this though, is surprising. Either Google has some monstrous server farm somewhere, or they're counting on content "owners" not using this utility too much that their processing queue becomes backed up.

Remember that it's not just the initial analysis/data extraction to some form of meta-data representation (eigenvectors or wavelet data) that has to be performed. Every subsequent video submission by every teenager out there has to be run through the same video analysis process and then compared to the entire library.

Re:Circumvention Ideas

[ZorbaTHut]

I've actually written a video comparison utility, and it would have neatly ignored every single one of these (with the exception of "backwards", which would have taken about five more minutes of work - it wasn't really important in my case.) Video is an interesting case because it's already so damaged by the very nature of compression, your tester has to be very lax to catch anything - but on the other hand, there's so much data that it's easier than you'd think to match up. Especially if you're willing to toss borderline cases at human checkers - you honestly end up with surprisingly few of those.

I don't know what Google is doing along these lines, though.

Another Site With Automated Content Filtering

[szyzyg]

You might have missed out on imeem.com [imeem.com] or at least ignored them ever since they changed from being a client/IM based p2p network to being a social media site about 2 years ago. But for the last 6 months they've been using automated content filtering for the music that people are posting to the site. Some of the people who register their content are have deals with imeem which allows the free sharing of their music - labels like Warners, Sony, BMG, Nettwerk, Beggars etc etc, and of course there are a few labels who have their tracks reduced to 30 second samples.

It should be noted that imeem announced all its big deals after turning its system on so presumably the content identification system helped make those media deals possible.

Re:Circumvention Ideas

[Applekid]

Yes, but given the quality of the previous fingerprinting, all those tricks are likely to work.

One video my, er, friend was uploading (that's my story and I'm sticking to it) was removed from youtube. He tried uploading it again and it didn't even go up, it was just immediately rejected. Out comes the hex editor and he changed the last byte to something else and reuploaded. It worked like a peach, like they were just doing checksums on the upload. *rollseyes*

For how long their fingerprinting has been in the making, one can only hope it's as functional as your comparison utility.

Add my vote for:
a1) chroma-shifting during encode
a2) video rotated 180 degrees, to be corrected with nvidia's nview "rotate monitor"
a3) odd, non-standard framerates (27 fps, etc)

Re:Circumvention Ideas

[PeterBrett]

How about subtly shifting each pixel one pixel in a random direction (ensuring that they all end up heading in the same direction for any particular frame) and then making each pixel a slightly different color shade, you'd have to accept a good number of false positives to be able to catch videos in a different location with different colors than the original.

Dead easy to spot. Ever heard of sift descriptors? They're fast to compute, and you only need one or two per frame to be able to uniquely fingerprint a video in a way that's totally resistant to rotation, recolouring, frame rate changes, and most of the other (lame) circumvention techniques suggested in this discussion.

Re:Circumvention Ideas

[Animaether]

a1) ignore chroma data (gets compressed more anyway), or compare relative (rather than absolute) values - done
a2) to fall in line with 'use custom player to ungarble garbled content'; users don't want to have to jump through hoops to play back videos. Btw, are you going to rotate the audio, too? - done
a3) base your fingerprint on the realtime performance, not on exact frames. Use a margin of, say, +-5%. Anything over that will result in a 'garbled' up video again anyway.

In essence it comes down to this... if you take any decent fingerprinting software, then the only reasonable way to get around them is by garbling the video; at which point people don't want to watch it anymore, or would have to jump through hoops to get a special player to ungarble. 'Mission accomplished' for the content copyright holders.

It's funny that anytime this sort of thing pops up, most people are heavily debating how to defeat the system, rather than worrying about their own original content (or parody content/etc.) getting falsely flagged.

Re:Opt Out!?

[RobertM1968]

Not just that, but it is going beyond what the DMCA is requiring (by making the takedown request method easier than required).

There are additional implications (as recently reported on /.) which I think will be worsened by this... for instance, a Viacom or an RIAA "clicking" takedown requests on a lot more content (that isnt theirs) now that it is much easier to do so. This is already a growing problem - I predict it will just worsen now that it is even easier for them.