Multimodal deepfake detection and localization based on FPN-Transformer

The researchers propose a multimodal deepfake detection and localization framework based on FPN-Transformer, achieving a score of 0.7535 on the IJCAI'25 DDL-AV benchmark. The framework uses WavLM and CLIP to extract hierarchical temporal features from audio and video, and builds a multi-scale feature pyramid through R-TLM blocks with local attention. A dual-branch prediction head simultaneously predicts the forgery probability and regresses the temporal offset of tampered segments, achieving frame-level localization accuracy.