Architecture-Adaptive Uncertainty Fusion for Deepfake Detection

2026-06-04

AI Summary

This work addresses the lack of reliable quantification of prediction uncertainty in existing deepfake detection methods, which hinders their adaptability across diverse model architectures and real-world forensic scenarios. The authors propose COF, an architecture-agnostic uncertainty fusion framework that requires no model modification. By formulating a constrained optimization problem on the probability simplex, COF linearly combines five complementary uncertainty sources—epistemic, aleatoric, calibration-based, conformal, and distributional—and learns optimal fusion weights in just 42 seconds. Evaluated on FaceForensics++, COF demonstrates strong cross-architecture generalization and significantly outperforms random forest baselines on nine out of eleven mainstream architectures in cross-domain tests on CelebDF. It achieves up to a 7.3× improvement in the correlation between uncertainty and prediction error while exhibiting notably lower performance degradation across domains, thus offering a robust and efficient solution for practical deepfake detection.

Abstract

Deepfake detection systems achieve near-perfect accuracy on benchmarks, yet forensic deployment demands reliable prediction uncertainty. Existing uncertainty quantification (UQ) methods rely on single sources and ignore that optimal uncertainty composition varies across architectures. We propose Correlation-Optimized Fusion (COF), an architecture-adaptive framework that 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. COF requires no model modifications and only 42 s of weight optimization, compared to 20--45 h for a 5-model Deep Ensemble. Evaluation across eleven architectures on FaceForensics++ reveals a fundamental trade-off: under matched train/evaluation protocol, non-linear methods achieve approximately 5--6% higher in-domain correlation than COF (mean r = 0.438), but this reverses under distribution shift. On CelebDF, COF outperforms Random Forest in 9/11 architectures with up to 7.3x higher correlation (MaxViT-B: r = 0.249 vs. 0.034); RF degrades 85% cross-domain to r = 0.071, whereas COF retains substantially more signal (74% drop to r = 0.116). Cross-dataset evaluation on CelebDF and DFDC reveals catastrophic generalization failure across all methods: in-domain correlations of 0.41--0.47 collapse to near-zero externally (mean degradation 90.7%), with seven of eleven architectures exhibiting uncertainty inversion. These results establish COF as a practical, interpretable framework for controlled-distribution deployment and identify domain-adaptive UQ as the central open challenge for forensic deployment.

Problem

• Deepfake Detection
• Uncertainty Quantification
• Distribution Shift
• Architecture Adaptation
• Forensic Deployment

Innovation

• Correlation-Optimized Fusion
• Uncertainty Quantification
• Deepfake Detection
• Architecture-Adaptive
• Distribution Shift