Computer scientists at UC Riverside have unveiled a new method that helps detect deepfake videos with greater accuracy.
A deepfake is a super-realistic image, video or audio clip that is not real. Deepfakes are not created by humans, but automatically created by artificial intelligence software programs.
The method is called Expression Manipulation Detection or EMD. This method divides the task of identifying a deepfake into two components.
The first part deals with discerning facial expressions and feeding information about the regions that contain the expressions such as the mouth, eyes or forehead. The second part is known as the encoder-decoder architecture and is responsible for manipulation detection and localization.
Identifying faces where only the expressions have been changed is difficult. The EMD method can identify and locate the specific regions within an image that have been changed.
The use of EMD to verify two sophisticated data sets for facial manipulation shows that the method performed better not only in detecting facial expression manipulation, but also identity swaps.
The sources for this piece include an article in UCR.