The proliferation of deepfakes has increasingly undermined confidence in the authenticity of online content. Given the rapid development of deepfake generation algorithms, new fake categories may keep appearing, posing a major challenge to existing deepfake detection methods. Despite recent progress in the development of deepfake detection methods for unseen forgeries, they are limited to binary real-vs-fake classification and can not detect the appearance of a novel category of fakes.
In this work, we study the Open Set DeepFake Detection (OSDFD) problem, which further demands that the detection model recognize novel fake categories instead of simply distinguishing real from fake. We reformulate the OSDFD problem, making it more applicable in real life. Then, we propose the DLED approach, which collects and fuses category-specific evidence in spatial and frequency levels.
Extensive evaluations across diverse settings demonstrate that the proposed DLED method achieves state‐of‐the‐art performance in recognizing forgeries originating from the novel category and exhibits competitive performance in real-versus-fake deepfake detection.
