Hacker News Reader

Comment by givinguflac | Hacker News Reader

This bothers me too. Sure one can claim there's a legal remedy, but Youtube was built on pirated content, and Google knows it. Even today I can go watch many, many pirated shows on youtube. It's been years that uploaders have evaded detection by simply mirroring (left to right, not hosting elsewhere as well) the video or slightly speeding up/slowing down the audio and/or video. Either google is willfully ignorant or they aren't as great at 'searching' for pirated content as they'd like us to believe. I think the former is much more likely.

View on HN

> Either google is willfully ignorant or they aren't as great at 'searching' for pirated content as they'd like us to believe. I think the former is much more likely.

Disclaimer: I work for Google (but not YouTube). Opinions are my own.

I suspect this has to do with economics. YouTube has way too many videos so everything is probably largely automated. Having to go through videos and find all possible modifications to pirated content is really difficult, and it'd be very expensive for them to do both in terms of engineering effort and also computing resources.

After all, it's not like you can search videos by their content, you can only search their titles or descriptions with text. Sure, they could start detecting things that have mirrored content, but then people would just change the method by which they modify the original video. People who are trying to upload pirated content only have to modify the videos they upload, while Google has to check all videos that are uploaded, the majority which are likely not pirated, and checking the videos is likely much more computationally expensive than modifying them a tiny bit.

I admit I know nothing at all about video formats and that stuff so if I'm wrong please correct me as I'm always grateful to be able to learn more.

it's not like you can search videos by their content

Err... yes you can. It's trivial.

Care to explain how you would do that?

I would construct a list of invariances that a classifier should be robust to, then construct (either empirically, analytically or via a machine learning algorithm using a synthetic dataset) a series of features that are invariant under these transformations. This is sometimes called "fingerprinting" and Shazam uses something like it to look up audio. I'm sure there are many specific approaches described in the literature.

I would then use these features to identify videos that have the same content as these already reported - pick your favorite clustering algorithm. This would ensure that for a single copyrighted video being reported all plausible manipulations that fall under the model are also discovered.

Point being, this is something that could be implemented in less than a month or two by an individual or small team. If Google hasn't implemented it it's because they don't want to, not because it can't be done.

That's a particularly robust model of solution. Even simpler solutions could match subsets, eg. normalized/simplified audio waveforms (cheap) or image stills (eg. using imagemagick) at probable change-of-scene points (eg. MP4 keyframes). Many people here could hack together something to do this in under one day.

True! And one should always KISS to start, at least.

Wow thank you for the detailed response. Seems like you are much smarter than I am!

How do you know this would perform any better than what they are currently doing though? Do you think this would catch the cases which currently pass their filtering like videos with mirrored images? And would this produce less false positives?

According to this post, it seems they already do something similar to what you've described: https://www.quora.com/Does-Content-ID-look-for-a-match-of-th...

I suppose in the end this is all going to just be speculation since they haven't disclosed what they actually do to try to find copyrighted content. Can you even predict how well a machine learning or classification algorithm will do without just trying it empirically?

What you can say is if there are N videos and k types of transforms that are used to "hide" content this would be hypothetically perfect with O(n + k) train examples - naive "reporting only" would be O(nk). This means that you should just need one example of each transform and one known instance of a particular copyright infringing media in order to get high performance.

Mirroring etc. would just be one example of the transforms an approach like this would be robust to.

It's hard to say what the error rates will be, but you should be on the last 5-10% Pd/Pfa pretty much instantly and then if you want more you can just acquire more true labelled data (if a user can find a copyrighted video on your site, I would think a contractor could, too!).

This well beyond what I understand so excuse me if I'm misunderstanding but doesn't this imply that they'd need to guess what k transforms might be used? This just seems like a cat and mouse type ordeal which I can only assume is something they would like to avoid as the complexity of and costs of running the system only gets worse, while the complexity does not add up on the side of those making pirated videos.

You're doing your employer a disservice :) Google implemented a system called Content ID and does check every video that is uploaded.

Oftentimes uploaders have to heavily crop or otherwise distort the video to get it to upload without getting caught by the detection system.

Haha oops. But thanks for the reference.

I figured they were already checking the videos somehow, but it seems to not catch videos that are mirror imaged or have other modifications. What I was trying to say was that it's much easier and cheaper for videos to be modified than it is to add additional filters through which uploaded content must be screened for copyrighted material.

I would say it's more so the case with Pinterest. However, YouTube does flag copyrighted content and in many cases the copyright holder is collecting video ad revenue even though somebody illegally uploaded it. This is almost always the case with music.

pinterest? people don't go to pinterest for the express reason of watching copyrighted material

You seem to assume that pictures are not copyrighted material.

Well if you payed attention to my wording you would have noticed that I was talking about the intention of the people visiting the website. Also people don't make money from uploading pictures to pinterest.

There is no financial or business incentive for Google to block pirated content.

Reply on news.ycombinator.com