An Analysis Of YouTube's Content Fingerprinting System
from the data-data-data dept
There have been plenty of questions about YouTube's fingerprinting system and how it works, and the EFF notes that an enterprising YouTube user figured that with a little experimentation, he could perhaps figure out how the system worked. Basically, he uploaded 82 different versions of a song, to see what YouTube caught... and what it didn't. He's put together a list of what he found out that's worth reading through, noting that it could still "catch" plenty of distorted content and that it reviews every video as soon as it's uploaded. But perhaps the most interesting (and surprising) bit: it seems to only look at the beginning of the video. He found that if he left the first 30 seconds blank, the system didn't catch anything. But if just played the first 30 seconds of a song, the video got flagged... There are some other interesting findings as well, that pretty much highlight how questionable some of these fingerprinting systems are in terms of accurately identifying the content in question.