ER-FusNet: RGB–주파수 이중 스트림의 MLP 후기 융합을 이용한 딥페이크 탐지

초록

To address generalization degradation from domain shifts in deepfake detection, we propose ER-FusNet, a dual-stream model that fuses RGB and DCT-based frequency cues via cross-attention. ER-FusNet couples an RGB EfficientNet with a DCT-based RepLKNet and, in cross-dataset tests, achieves F1 scores of 0.92–0.99 on benchmarks like Celeb-DF v2 while raising WildDeepfake F1 from 0.53 to 0.65 (~12% over the single-stream average). These results show that jointly leveraging global RGB signals and fine-grained frequency artifacts yields robust real-world deepfake detection.

키워드