Adaptive Gated Deepfake Detection for Low-Resolution and Resource-Constrained Environments

The researchers proposed an adaptive gating deepfake detection framework called AdaGate-DF, aimed at solving the detection challenges in low-resolution images and scenarios with limited computing resources. This framework uses image quality cues to route samples to a dual-multi-exit system, allowing high-quality inputs to produce results in advance to save computational resources. Tests on the Celeb-DF benchmark dataset showed that the framework achieved an AUC of 0.9370, outperforming MaD-CoRN and DefakeHop++; when the input resolution was increased to 384 by 384, the AUC reached 0.9708. Additionally, the model performed well on the FaceForensics++ dataset, effectively addressing class imbalance issues and achieving a balance between detection performance, uncertainty prediction, and computational efficiency.