Detection of Deepfake Videos Using Long Distance Attention

Abstract

Recent advances in deepfake methods have made face video forgery a real possibility, capable of producing extremely misleading video footage and posing serious security risks. Furthermore, identifying these fake videos is a very difficult and time-sensitive task. Current detection algorithms typically approach the problem with a straightforward binary classification mindset. Since real and synthetic faces differ in minute ways, this study treats the topic as a novel fine-grained classification problem. Existing face forgery systems typically leave behind certain common artifacts in the spatial and temporal domains. Inconsistencies between frames in the time domain and generating errors in the space domain were among these objects. Furthermore, we introduce a spatial-temporal model that includes both the location and the passage of time in its representation of forging traces from a worldwide perspective. In creating both components, a novel long-distance attention technique was employed. The spatial domain component allows for a single frame capture of an artifact, while the time domain component allows for sequential frames. They make attention maps that look like patches. A more holistic view is provided by the attention approach, which aids in the extraction of local statistical information and the better assembly of global information. Lastly, similar to other granular classification techniques, the network is directed to concentrate on critical areas of the face by means of attention maps. Experimental results on several public datasets show that the suggested method outperforms the state-of-the-art, and the suggested longdistance attention strategy successfully captures crucial components for face forgeries