Multi-Cue Temporal Artifact Modeling for Deepfake Video Detection

Abstract

Face swapping, facial reenactment, and expression editing have extended deepfake traces from single-frame appearance to cross-frame dynamics. Existing detectors often depend on one spatial representation or temporal path, limiting the joint preservation of transferable semantics and extraction of fine-grained temporal anomalies. We propose Multi-cue Temporal Enhancement with CLIP (MTE-CLIP), a video-level detector with a frozen CLIP image encoder. Its side decoder combines facial-prior spatial interaction and rotary temporal context with continuous-frame correlation, local temporal inconsistency, and token-domain temporal-frequency cues. Text-derived prototypes provide stable semantic anchors, and fixed residual addition integrates the three cues. Spatial, temporal, and joint-head probabilities are averaged for prediction. Experiments on FF++ and six external datasets show competitive performance across manipulation types and target domains; ablations confirm complementary but dataset-dependent cue effects.