Adaptive routing based on input quality lets deepfake detectors achieve high accuracy while reducing compute costs—a practical approach for deploying detection on phones and edge devices where both accuracy and speed matter.
This paper presents AdaGate-DF, a deepfake detection system that adapts to image quality by routing samples through different computational paths—high-quality images exit early to save compute, while low-quality images get more processing. It achieves strong detection accuracy (AUC 0.9370 on Celeb-DF) while maintaining low inference latency, making it practical for resource-constrained devices.