AdaGate-DF Routes Deepfake Samples by Image Quality So Clean Frames Exit Early
arXiv 2609.05320·low signal
arXiv 2609.05320 targets deepfake detection under low resolution and constrained compute, where fixed inference paths waste work on easy samples. AdaGate-DF uses image-quality cues to route samples through a dual multi-exit system, letting high-quality images exit earlier. Against MaD-CoRN, DefakeHop++ and ShuffleNetV2 on Celeb-DF and FaceForensics++, it reached AUC 0.9370 on Celeb-DF with low inference latency, rising to 0.9708 as input resolution increased.