Copyright detection in videos based on channel context
The herein disclosed technology provides methods and systems that utilize machine learning solutions to identify web-based channels that are sources pirated copyright material, such as by using a machine learning classifier that is trained on historical copyright piracy data and channel features that may be determined and analyzed for each of a collection of channels without analyzing video or audio content of the channel.
1. A method of identifying a channel as being a source of pirated copyrighted material, the method comprising:
inputting a set of channel features to a feature-based classifier trained based on a training dataset including channel features and historical copyright piracy data for each of multiple channels, the historical copyright piracy data included in the training dataset including a binary indicator for each of the multiple channels that indicates whether the associated channel is preidentified as a source of pirated copyright material, each one of the channel features being a channel attribute determined without analyzing video or audio content of the channel;
receiving an output from the feature-based classifier, the output including a first copyright violation prediction indicating a probability that copyright-protected material is available through the channel; and
executing a remedial action when the first copyright violation prediction satisfies predefined criteria.
2. The method of claim 1 , wherein the channel features include one or more features from a group comprising:
a ratio between a number of subscribers to the channel and a number of active videos available on the channel;
a geographical source associated with the channel;
an age of the channel;
channel banner;
historical changes in channel viewership; and
metadata indicating ownership claims to content of the channel.
3. The method of claim 1 , wherein the channel features include channel metadata including at least one of a channel description and a title.
4. The method of claim 1 , further comprising:
determining that the copyright violation prediction output by the feature-based classifier exceeds a first predefined threshold but does not exceed a second predefined threshold;
analyzing video content of the channel responsive to the determination, the video content including at least one of image data and audio data;
rendering a second copyright violation prediction based on the analysis of the video content; and
executing a remedial action responsive to determining that the second copyright violation prediction satisfies predefined second criteria.
5. The method of claim 1 , wherein the remedial action triggers at least one of a take-down of content of the channel and a revocation of content-sharing permissions for an owner of the channel.
6. The method of claim 1 , wherein the remedial action creates a log entry flagging the channel for subsequent manual review.
7. The method of claim 1 , wherein the feature-based classifier implements a random forest model.
8. A system for identifying a channel as being a source of pirated copyrighted material, the system comprising:
a feature-based channel classifier stored in memory and executable to:
analyze a received set of channel features for the channel, each one of the channel features being a channel attribute determined without analyzing video or audio content of the channel; and
output a first copyright violation prediction indicating a probability that copyright-protected material is available through the channel; and
a remedial action initiator stored in memory and executable to initiate a remedial action when the first copyright violation prediction satisfies predefined criteria, the feature-based channel classifier being a machine learning model trained on a training dataset including channel features and historical copyright piracy data for each of multiple channels, the historical copyright piracy data for each channel of the multiple channels indicating whether the channel is preidentified as a source of pirated copyright material.
9. The system of claim 8 , wherein the channel features include one or more features from a group comprising:
a ratio between a number of subscribers to the channel and a number of active videos available on the channel;
a geographical source associated with the channel;
an age of the channel;
channel banner;
historical changes in channel viewership; and
metadata indicating ownership claims to content of the channel.
10. The system of claim 8 , wherein the channel features include channel metadata including at least one of a channel description and a title.
11. The system of claim 8 , further comprising:
a content-based classifier stored in memory and executable to:
analyzing video content of the channel responsive to a determination that the copyright violation prediction output by the feature-based classifier exceeds a first predefined threshold but does not exceed a second predefined threshold, the video content including at least one of image data and audio data; and
render a second copyright violation prediction based on the analysis of the video content, wherein the remedial action initiator is further executable to initiate the remedial action responsive to determining that the second copyright violation prediction satisfies predefined second criteria.
12. The system of claim 8 , wherein the remedial action triggers at least one of a take-down of content of the channel and a revocation of content-sharing permissions for an owner of the channel.
13. The system of claim 8 , wherein the remedial action includes creating a log entry flagging the channel for subsequent manual review.
14. The system of claim 8 , wherein the feature-based classifier implements a random forest model.
15. One or more memory devices encoding instructions for executing a computer process comprising:
inputting a set of channel features to a feature-based classifier trained based on a training dataset including channel features and historical copyright piracy data for each of multiple channels, each one of the channel features being a channel attribute determined without analyzing video or audio content of the channel, the historical copyright piracy data included in the training dataset for the feature-based classifier includes a binary indicator for each of the multiple channels, the binary indicator indicating whether the associated channel is preidentified as a source of pirated copyright material;
receiving an output from the feature-based classifier, the output including a first copyright violation prediction indicating a probability that copyright-protected material is available through the channel; and
executing a remedial action when the first copyright violation prediction satisfies predefined criteria.
16. The one or more memory devices of claim 15 , wherein the channel features include one or more features from a group comprising:
a ratio between a number of subscribers to the channel and a number of active videos available on the channel;
a geographical source associated with the channel;
an age of the channel;
channel banner;
historical changes in channel viewership; and
metadata indicating ownership claims to content of the channel.
17. The one or more memory devices of claim 15 , wherein the computer process further comprises:
determining that the copyright violation prediction output by the feature-based classifier exceeds a first predefined threshold but does not exceed a second predefined threshold;
analyzing video content of the channel responsive to the determination, the video content including at least one of image data and audio data;
rendering a second copyright violation prediction based on the analysis of the video content; and
executing a remedial action responsive to determining that the second copyright violation prediction satisfies predefined second criteria.