Frequency perturbation regularization for cross-dataset deepfake detection
Abstract
Deepfake detectors trained on a single source dataset often lose accuracy when the test data come from different manipulation pipelines, compression settings, or acquisition conditions. This paper presents a cross-dataset deepfake detection method based on frequency perturbation regularization. The detector uses a CLIP-L/14 visual encoder with a lightweight LinearNorm classification head. During training, each input face image is transformed into the frequency domain, and mild random perturbations are injected into high-frequency components before reconstruction. This input-level regularization exposes the detector to more diverse frequency distributions and reduces its dependence on source-domain texture and compression artifacts. At inference time, a horizontal-flip test-time ensemble averages the logits of the original and flipped frames to improve video-level prediction stability. Experiments are trained on FaceForensics++ and evaluated on Celeb-DF v2, DFDC, and Celeb-DF v1. The proposed Freq+TTA configuration achieves video-level AUROC scores of 96.49%, 87.96%, and 95.48%, respectively, showing consistent gains over the CLIP-L/14 baseline. The results verify that simple frequency-domain regularization and lightweight inference enhancement can improve generalizable deepfake detection without introducing a complex extra branch.