When AUC Misleads: Polarization-Aware Evaluation of Deepfake Detectors under Domain Shift

Dat Nguyen, Cosmin Radoi, Romain Hermary, Marcella Astrid, Nesryne Mejri, Enjie Ghorbel, Djamila AouadaarXiv 2026 · arXiv

TL;DR

Standard evaluation protocols for deepfake detectors rely on averaging the Area Under the ROC Curve (AUC) across individual datasets, which dangerously overestimates real-world performance. This paper introduces Cross-dataset AUC (Cross-AUC), a novel metric that incorporates prediction polarization via Wasserstein Distance to accurately evaluate how well a detector handles real-world domain shifts and mixed data sources.

What Problem Does This Paper Address?

When researchers train deepfake detectors, they usually evaluate generalization by testing the model on several independent datasets, calculating the AUC for each, and averaging the results. However, this approach assumes that the model's predicted score distributions are homogeneous across different data sources.

In reality, real-world deployments present a messy mixture of data with varying compressions, lighting conditions, and manipulation techniques (domain shift). A detector might perfectly separate real from fake images within Dataset A (using an optimal threshold of 0.3) and within Dataset B (using an optimal threshold of 0.6). But when these datasets are mixed together, their unaligned score distributions overlap significantly. This causes a severe drop in actual performance that traditional average AUC completely hides.

Fig. 1: Example of optimal thresholds for predictions on different datasets, A (top) and B (middle). While a perfect AUC = 1 is reached on both datasets separately, an obvious drop in AUC will appear when combining them to simulate real-world settings (bottom), since the scores are no longer well ordered. The different nature of score distributions leads to a clear disparity between optimal thresholds (purple range), demonstrating the importance of taking into account the polarization of probability predictions in real-world scenarios.

Methods

To solve this, the authors propose Cross-dataset AUC (Cross-AUC). Instead of a simple average, Cross-AUC adjusts the performance metric by looking at prediction polarization—how confidently and cleanly a model separates real and fake score distributions across the board.

  1. Quantifying Polarization via Wasserstein Distance (WD): An ideal model pushes real samples to a score of 0 and fake samples to 1 (creating Dirac impulses at the boundaries). The authors use the Wasserstein Distance to measure the gap between the positive (fake) and negative (real) score distributions. A larger gap means higher polarization, indicating that a single, fixed threshold will remain robust even if domain shifts cause minor score drifts.
  2. The Cross-AUC Formulation: The metric applies the harmonic mean to individual dataset AUCs and combines it with a penalty-reward criterion. It penalizes models that exhibit high variance (instability) in AUC or polarization across datasets, while rewarding models that maintain highly polarized, well-separated score distributions.

Datasets

The evaluation framework was tested using a cross-dataset protocol where models were trained on one dataset and tested on others:

  • Training Source: FaceForensics++ (FF++)
  • Evaluation Benchmarks (7 datasets): Celeb-DF (CDF), Google Deepfake Detection (DFD), WildDeepfake (DFW), Deepfake Detection Challenge (DFDC), Deepfake Detection Challenge Preview (DFDCP), and DF40.
  • Out-of-Distribution (OOD) Testing: MagicBrush (a non-facial image editing dataset used to verify the metric's generalizability to broader binary classification tasks).

Key Insights and Ideas

  • The "Perfect AUC" Illusion: A model can achieve a flawless AUC of 1.0 on multiple datasets individually but fail in practice if its decision boundaries are misaligned across domains.
  • Polarization Equals Robustness: High polarization creates a wide decision margin. This safety buffer ensures that threshold instability across different data sources won't immediately degrade classification accuracy.
  • Uncovering Rank Reversals: Traditional average AUC can rank models incorrectly. For instance, in the paper's experiments, ForensicAdapter achieved a higher average AUC than LAA-Net. However, LAA-Net actually performed better on a combined dataset because its score distributions were better aligned—a reality captured by Cross-AUC but missed by standard averaging.
Fig. 4: Example of the obtained optimal thresholds on two datasets, A (top) and B (middle). Although both detectors achieve perfect AUC = 1 when evaluated on each dataset separately, mixing A&B (bottom) can cause a sharp AUC drop (Detector X), while a more stable detector preserves ordering and maintains AUC = 1 (Detector Y). It can be observed that a larger gap exists between the two classes in the score space, reducing the impact of a varying threshold. This motivates adopting a polarization- aware evaluation.

Key Results

  • Average AUC Overestimates Performance: Across all tested state-of-the-art models (including Xception, SLADD, RECCE, SBI, CADDM, LAA-Net, and ForensicAdapter), the average AUC was consistently higher than the true combined dataset AUC ($\text{AUC}_c$), with overestimating gaps reaching up to 12.3 percentage points.
  • Cross-AUC Tracks Reality: Cross-AUC closely mirrored the actual combined dataset AUC ($\text{AUC}_c$) without requiring researchers to physically pool the datasets together. For example, for LAA-Net, the true combined AUC was 86.51%, and Cross-AUC estimated it at 86.47% (a difference of just 0.04%).
  • Robustness Across Estimators: The metric proved stable whether using Wasserstein Distance or KL Divergence as the underlying polarization estimator, though Wasserstein Distance provided the most stable results under severe distribution shifts.

Limitations and Future Work

  • Lack of Theoretical Grounding: While Cross-AUC is highly practical and empirically validated, the exact mathematical formulation of its penalty-reward function lacks rigorous theoretical proofs.
  • Application Scope: The authors focused exclusively on deepfake detection for this study. However, because the vulnerability of AUC to domain shift is a fundamental property of binary classification, future work will involve validating Cross-AUC across other machine learning domains facing distribution shifts.

Summary updated

Abstract

Recent advances in generative AI, such as diffusion models and face-swapping tools, have enabled the creation of highly realistic deepfakes, leading to real-world harms including financial fraud and non-consensual explicit content. In response, deepfake detection has become an active research area, with recent methods increasingly focusing on improving generalization to unseen manipulations. This is typically evaluated using the Area Under the ROC Curve (AUC) measured separately across multiple datasets. However, such an evaluation fails to reflect real-world scenarios where detectors face a mixture of data sources and varying artifact types. To address this limitation, we introduce a novel metric, Cross-dataset AUC (Cross-AUC) that averages per-domain AUCs with a measure of prediction polarization for taking into account the robustness to domain shift. The polarization extent is quantified by the Wasserstein Distance between class score distributions. Cross-AUC not only assesses the generalization capabilities of deepfake detectors under domain shifts more realistically, but it is also interpretable as it better explains the reason behind a drop in performance. Experiments performed on seven benchmark datasets demonstrate its practical relevance.
BibTeX
@misc{nguyen2026when,
  author = {Dat Nguyen and Cosmin Radoi and Romain Hermary and Marcella Astrid and Nesryne Mejri and Enjie Ghorbel and Djamila Aouada},
  title = {{When AUC Misleads: Polarization-Aware Evaluation of Deepfake Detectors under Domain Shift}},
  year = {2026},
  eprint = {2606.19184},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.19184},
}
Figure