Deepfake Detector Generalization Under Fake-Generation Evolution: A Controlled Distribution-Shift Study

This paper investigates the generalization of deepfake detectors under controlled fake-generation distribution shift. Two paired datasets derived from FaceForensics++ are used: a benchmark-era dataset using Deepfakes face swapping and a diffusion-era dataset using Stable Diffusion face-region inpainting, while holding the real-image distribution constant. Xception and DINOv2-based detectors are evaluated using a 2×2 train/test protocol to measure both in-domain and cross-domain performance. Results show strong in-domain performance but substantial degradation under cross-domain evaluation, demonstrating that benchmark-only evaluation can overestimate robustness to evolving manipulation methods.