Learning Face Forgery Detection in Unseen Domain with Generalization Deepfake Detector

Abstract

Face forgery generation algorithms have advanced rapidly, resulting in a diverse range of manipulated videos and images which are difficult to identify. As a result, face manipulation using deepfake technique has a significantly increased societal anxiety and posed serious security problems. Recently, a variety of deep fake detection techniques have been presented. Convolutional neural networks (CNN) architecture are used for most of the deepfake detection models as binary classification problems. These methods usually achieve very good accuracy for specific dataset. However, when evaluated across datasets, the performance of these approaches drastically declines. In this paper, we propose a face forgery detection method to increase the generalization of the model, named Generalization Deepfake Detector (GDD). The Generalization Deepfake Detector model has ability to instantly solve new unseen domains without the requirement for model updates.