COF Framework Boosts Deepfake Uncertainty but Fails Cross-Dataset

A new paper from Ritesh Sharma and colleagues introduces Correlation-Optimized Fusion (COF), an architecture-adaptive framework for uncertainty quantification in deepfake detection. COF fuses five complementary uncertainty sources—epistemic, aleatoric, calibration, conformal, and distributional—by maximizing Pearson correlation between fused uncertainty scores and prediction errors via constrained optimization on the probability simplex. The method requires no model modifications and completes weight optimization in just 42 seconds, a dramatic improvement over the 20–45 hours needed for a 5-model Deep Ensemble. Evaluated on FaceForensics++ across 11 architectures, COF achieves strong in-domain performance: on CelebDF, it outperforms Random Forest in 9/11 architectures with up to 7.3x higher correlation (MaxViT-B: r = 0.249 vs. 0.034).

However, the paper also reveals a critical limitation. Under matched train/evaluation protocols, non-linear methods achieve approximately 5–6% higher in-domain correlation than COF (mean r = 0.438), but this reverses under distribution shift. Cross-dataset evaluation on CelebDF and DFDC shows catastrophic generalization failure across all methods. In-domain correlations of 0.41–0.47 collapse to near-zero externally, with a mean degradation of 90.7%. Seven of the eleven architectures exhibit uncertainty inversion, where the model becomes confidently wrong. These results establish COF as a practical, interpretable framework for controlled-distribution deployment but identify domain-adaptive UQ as the central open challenge for real-world forensic deepfake detection.