PlainProbe: A Stable Cross-Entropy Baseline for Data-Scarce Deepfake Detection

Deepfake detectors are difficult to stabilize under new generators and source distributions. We present an objective-ablation framework for adapting deepfake detectors with limited data. Using a fixed GenD-based CLIP detector, we reformulate training as standard cross-entropy over real and fake labels by removing auxiliary alignment and uniformity objectives. Our empirical results show that this simplified objective yields more stable training, reducing AUROC standard deviation across five random seeds by 57--75\% relative to a matched GenD-based reference. On Celeb-DF v2, FFTEST, and WILDD, our method achieves AUROC/mAP scores of 0.7287/0.7000, 0.9867/0.9857, and 0.8410/0.8457, respectively. These results suggest that stable, reproducible objectives can serve as useful baselines for reliability-oriented deepfake detection research.