Google introduced the content ID system in 2007 to cope with the unauthorized use of copyrighted music and video content. The content ID system has successfully helped the copyright owners to find the pirated content and paid billions of dollars to the copyright owners1.
The copyright owners can submit the audio and video content they want to protect to YouTube. When a user uploads new content to YouTube, the new content will be compared to the files in the database to check their similarity.
How do we compare the similarity between audio or video files? We need audio and video fingerprinting technology. Generally speaking, we can extract fingerprints (i.e., features or vectors) from the audio and video files and store them in the database. When a new content arrives, we extract its fingerprints and compare it with the fingerprints in the database. Then we can calculate the similarity between the new content and the content in the database. If the similarity is above a threshold, we can assume that the uploaded content is a pirated content.
There is little resource online talking about which methods Google use for fingerprinting audio and videos.
The YouTube audio fingerprinting technology is quite advanced. As pointed out by videos here and here, it can spots audio piracy even if you distort the music significantly (by adding noise or change speed etc.).
This paper from Google research shows a method for audio fingerprinting. It may be used by content ID. However, this paper is quite old. YouTube may have improved its algorithms over the years.
For video fingerprinting, we would guess that YouTube extract frames from videos and use deep-neural nets to extract features from video frames. Those features are the so-called fingerprints for the video. The similarity between videos can be compared frame by frame and aggregated in some way, for example, by calculating how many similar frames two videos share in common.
This above guess is partially verified by post here. Apart from that, I can not seem to find resource on what methods YouTube actually use for video fingerprinting.
Since fingerprinting uses machine learning methods for finding pirated content, we can forge adversarial content to fool the content ID system: the pirated content can be altered in a subtle way so that the system can not detect it is a pirated content.
There are some research papers on this, for example, see here and here.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。