Only Train Once: Uncertainty-Aware One-Class Learning for Face Authenticity Detection

arXiv:2605.10040 · cs.CV · Submitted 2026-05-11 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Only Train Once".

Jane: Here's the gist of the paper. The gist The rapid evolution of generative paradigms has enabled the creation of highly realistic imagery, which escalating the risks of identity fraud and the dissemination of disinformation.<ref:2605.10040#pg2> Problem Formulation The paper reformulates face forgery detection as a one-class classification (OCC) task, treating manipulated faces as anomalous samples<ref:2605.10040#pg3> by shifting the objective from detecting forgeries to Face Authenticity Detection<ref:2605.10040#pg4>.

Tom: The formal goal is to learn a mapping function f: I → (y, u) ∈

R, 1: ×

R, 1: , subject to constraints that real images yield low uncertainty and forged images yield high uncertainty<ref:2605.10040#pg4>.

Jane: This approach allows the framework to generalize across diverse forgery methods by learning representations exclusively from pristine facial imagery<ref:2605.10040#pg5>.

Tom: FADNet Architecture The proposed FADNet architecture is designed to capture the intrinsic distribution of real facial data<ref:2605.10040#pg4>.

Title and authors: Tom: We've been talking about how this paper, "Only Train Once: Uncertainty-Aware One-Class Learning for Face Authenticity Detection," tackles the problem of generalizing face forgery detection by focusing on one-class learning and uncertainty modeling.

Jane: It’s a paper that fundamentally changes the goal from detecting specific fakes to learning the intrinsic distribution of what is real, using only authentic data to train the system.

Tom: Right, and they use a framework called FADNet, which reformulates face forgery detection as an one-class classification task where forgeries become anomalous samples.

Lu: The key idea is that by training exclusively on authentic facial data to capture their intrinsic representations, the model learns what genuine faces look like across diverse styles.

Meng: So, instead of constantly chasing new types of deepfakes, the model gets a solid base representation from the real faces it sees during training.

Jane: They use a method called Evidential Deep Learning to model this uncertainty through a Dirichlet distribution, which gives us that predictive uncertainty score u(I).

Tom: This uncertainty is then used to make the final call: if u(I) is low, it’s authentic; if it’s high, they flag it as likely forged.

Lu: They are using a multi-objective loss function with Evidential Loss, Uncertainty Regularization, and Feature Compactness Loss to guide this learning process.

Meng: The feature compactness part specifically helps keep the feature representation of real faces tightly clustered together in the embedding space.

Jane: So what does this mean for us? It means we can build a system that is less likely to break when it encounters a forgery technique we haven't seen before because it’s focused on the fundamental structure of authenticity.

Tom: The results are pretty solid, showing ninety-six point six three percent accuracy and ninety-eight point eight three percent precision on the DF40 dataset, which is significantly outperforming current state-of-the-art methods.

Lu: They also showed good generalization against newer synthesis paradigms, which is a big deal because it means their representation learning isn't brittle.

Meng: From an engineering standpoint, the fact that it reduces inference time by about ninety-one point eight percent compared to Idcnet on the DF40 dataset makes this practical for real-world deployment.

Jane: So, if you’re driving and you see a suspicious face online, this research suggests we can rely on a score based on uncertainty rather than just a binary detection.

Tom: We'll talk about how they improve on this approach in the next segment, specifically looking at those added components that make it even more robust.

The paper's summary: Tom: So we’ve established that FADNet works by treating face forgery detection as an OCC task and using uncertainty modeling to separate real from fake data.

Jane: Now let’s unpack the mechanism a bit more—how they actually build this system, especially with those losses they mentioned in the title.

Lu: They start by using a pre-trained ResNet-fifty backbone for feature extraction, which gives them a two thousand forty-eight-dimensional raw feature vector fraw.

Meng: That’s a standard starting point, but then they project that vector into a discriminative embedding space using a two-layer MultiLayer Perceptron, or MLP.

Tom: After the MLP, they feed that feature vector f into their evidential layer which parameterizes the concentration parameters of a Dirichlet distribution.

Jane: This is where the uncertainty u(I) comes from, and they define alpha as e plus one to control how that distribution behaves.

Lu: The predictive uncertainty u(I) is derived directly from this Dirichlet distribution, and a higher value means a higher degree of vacuity in the image.

Meng: During training, they introduce pseudo forgery samples generated by a PFIG to act as out-of-distribution data during the learning phase.

Tom: The loss function Ltotal combines the Evidential Loss for evidence accumulation, Uncertainty Regularization to control uncertainty distributions, and Feature Compactness Loss for real faces.

Jane: The Feature Compactness Loss specifically minimizes intra-class variance for real faces by constraining their feature distribution to form a cohesive cluster.

Lu: This constraint on the feature distribution is what they claim helps them effectively characterize the intrinsic distribution of authentic facial features.

Meng: If that clustering works well, it should translate into a much cleaner separation when we finally run inference in production environments.

Jane: So, the summary is: they learn a representation from real faces, measure how much an input image deviates from that learned structure using Dirichlet uncertainty, and use specific losses to make sure the real features are tight and forged ones are spread out.

Tom: That’s the high-level picture of what they accomplished in "Only Train Once: Uncertainty-Aware One-Class Learning for Face Authenticity Detection."

The paper's improvements: Tom: Now we look at what the paper suggests as ways to make this system even better, because it’s not just about the initial setup.

Jane: They suggest a few key improvements, starting with incorporating pseudo-forgery feature simulation during training to create more robust OOD data.

Lu: They suggest tightening decision boundaries around authentic data by applying four types of perturbations to simulate forgery artifacts, including Spectral Disruptions, Local Blending, and Checkerboard Artifacts.

Meng: So they are actively simulating different kinds of known artifacts so the model learns to be even more resilient against them during training.

Tom: Then they suggest using the evidential deep learning approach for calibrated uncertainty scores which models uncertainty via the Dirichlet distribution.

Jane: This allows us to assign higher uncertainty specifically to out-of-distribution fake images, creating a clearer separation between real and forged content.

Lu: They also suggest using an uncertainty-based thresholding strategy during inference, where authentic face images yield low values and others generate high values.

Tom: They found that setting this threshold at zero point six nine one yields the optimal accuracy at that point for classification.

Meng: That threshold is a concrete number, which means we know exactly what level of uncertainty to look for when deploying the model on real traffic.

Jane: And they also emphasize the feature compactness loss again, defining it as Lcompact = (f(I(i)r) - fr) squared to minimize intra-class variance for real faces <ref:2605.10040#pg2>.

Tom: If you look at that specific constraint, it’s what helps in enhancing their separability from forgeries, and ablation studies show that dropping this loss leads to a decrease in accuracy to seventy-six percent and an F1-score of eighty-six point eight seven percent.

Lu: So they’re showing that the combination of uncertainty regularization and feature compactness really refines the detection precision.

Conclusion: Tom: To wrap this up, "Only Train Once: Uncertainty-Aware One-Class Learning for Face Authenticity Detection" shows that one-class learning combined with uncertainty modeling is a viable way to handle the massive challenge of universal face forgery detection.

Jane: It gives us a reliable system where we rely on a calculated uncertainty score, and they’ve shown good calibration against existing baselines.

Lu: The synergy between the uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features in a very stable way.

Meng: For production, this means less maintenance because it’s more robust against novel attacks, and the reduced inference time is a plus for scaling up.

Jane: It gives us confidence that we can build systems that are better at spotting subtle anomalies without needing to train on every single possible fake technique.

Tom: It’s a solid approach based on training only once on real data and using uncertainty modeling to provide interpretability for digital forensics applications.

Lu: Ultimately, the paper demonstrates superior stability in calibration performance compared to existing baselines when combining these two components together.

Qingchao Jiang, Zhenxuan Hou, Zhiying Zhu, Zhenxing Qian, Xinpeng Zhang, Zaiwang Gu

East China University of Science and Technology · Fudan University · Agency for Science, Technology and Research (A*STAR)

cs.CV

Submitted: 2026-05-11

Updated: 2026-10-20

Comments: Accepted for publication in IEEE Transactions on Multimedia (TMM)

Code: https://github.com/hzx111621/Only-Train-Once

Project page: https://peterwang512.github.io/CNNDetection/4https://github.com/ZhendongWang6/DIRE5https://github.com/wangzhiyuan120/idcnet6https://github.com/LoveSiameseCat/MoE-FFD7https://github.com/Ekko-zn/Diff-based-AIGD

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: The gist The rapid evolution of generative paradigms has enabled the creation of highly realistic imagery, which escalating the risks of identity fraud and the dissemination of

Key concepts

One-Class Classification (OCC): This is a task where the AI learns the characteristics of 'normal' data—in this case, real faces—without being explicitly taught what a forgery looks like. The model treats any input that deviates significantly from these learned real patterns as an anomaly or potential fake, allowing it to generalize across different forgery types.
Evidential Loss (LEDL): This loss function encourages the model to build strong evidence supporting the correct class (real face). It works by maximizing the concentration of a Dirichlet distribution around the true label, ensuring that real images yield high confidence and forged images show low confidence.
Uncertainty-Based Thresholding: Instead of a simple yes/no classification, FADNet uses the model's uncertainty score to make decisions. Real faces produce low uncertainty scores, while forged faces produce high scores. A threshold (like 0.691) is set to decide authenticity, providing a calibrated way to separate real from fake content.

Terminology

Accuracy — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Generalization — 1× in this paper · explained in 13 episodes
Generalization Capabilities — 1× in this paper · explained in 2 episodes
Generator — 1× in this paper · explained in 2 episodes

Summary

The gist The rapid evolution of generative paradigms has enabled the creation of highly realistic imagery, which escalating the risks of identity fraud and the dissemination of disinformation.<ref:2605.10040#pg2>

Problem Formulation

The paper reformulates face forgery detection as a one-class classification (OCC) task, treating manipulated faces as anomalous samples<ref:2605.10040#pg3> by shifting the objective from detecting forgeries to Face Authenticity Detection<ref:2605.10040#pg4>. The formal goal is to learn a mapping function f: I → (y, u) ∈ [R, 1] × [R, 1], subject to constraints that real images yield low uncertainty and forged images yield high uncertainty<ref:2605.10040#pg4>. This approach allows the framework to generalize across diverse forgery methods by learning representations exclusively from pristine facial imagery<ref:2605.10040#pg5>.

FADNet Architecture

The proposed FADNet architecture is designed to capture the intrinsic distribution of real facial data<ref:2605.10040#pg4>. The backbone utilizes a pre-trained ResNet-50 for feature extraction, which yields a 2048-dimensional raw feature vector fraw<ref:2605.10040#pg5>. These features are projected into a discriminative embedding space using a two-layer MultiLayer Perceptron (MLP)<ref:2605.10040#pg5>. The resulting feature vector f is then processed by an evidential layer that parameterizes the concentration parameters of a Dirichlet distribution, where α = e + 1<ref:2605.10040#pg5>. The predictive uncertainty u(I) is derived from this Dirichlet distribution, with a higher value indicating a higher degree of vacuity<ref:2605.10040#pg5>.

Training Strategy and Components

The training phase utilizes a plug-and-play pseudo forgery image generator (PFIG) to create pseudo-forgery samples which act as OOD data during training<ref:2605.10040#pg5>. The model is trained using a multi-objective loss function Ltotal = LEDL + Lunc + λcLcompact<ref:2605.10040#pg6>.

Evidential Loss (LEDL) encourages the model to accumulate evidence αk for the correct class, maximizing the concentration of the Dirichlet distribution toward the ground-truth label<ref:2605.10040#pg6>.
Uncertainty Regularization (Lunc) explicitly constrains uncertainty distributions, penalizing high uncertainty in real images and suppressing evidence accumulation for forged images<ref:2605.10040#pg6>.
Feature Compactness Loss (Lcompact) minimizes intra-class variance for real faces by constraining their feature distribution to form a cohesive cluster<ref:2605.10040#pg6>.

Inference and Decision Making

During inference, the model produces Dirichlet parameters α(I) and the predictive uncertainty u(I)<ref:2605.10040#pg7>. A decision strategy based on an uncertainty-based thresholding is employed: y = 1 if u(I) ≤ τ, and y = 0 if u(I) > τ<ref:2605.10040#pg7>. The optimal decision boundary is found at a threshold τ = 0.691<ref:2605.10040#pg10>. This mechanism provides an intuitive reflection of the model’s confidence, where low uncertainty values are associated with authentic faces<ref:2605.10040#pg7>.

Experimental Validation

Extensive experimental evaluations on the DF40 and ASFD benchmarks demonstrate that FADNet achieves superior performance and generalization capabilities<ref:2605.10040#pg2>. The method substantially outperforms existing state-of-the-art (SOTA) methods, yielding a remarkable average accuracy of 96.63% and an average precision of 98.83% on the DF40 dataset<ref:2605.10040#pg8>. FADNet maintains good generalization capabilities against newest synthesis paradigms<ref:2605.10040#pg8>. Furthermore, the model provides calibrated uncertainty scores and significantly mitigates model overconfidence, enhancing the overall reliability of the detection process<ref:2605.10040#pg8>. The computational efficiency is also favorable, with FADNet reducing inference time by approximately 91.8% compared to Idcnet on the DF40 dataset<ref:2605.10040#pg8>.

Conclusion

FADNet is a one-class learning and uncertainty-aware method for face authenticity detection that leverages exclusively authentic facial data<ref:2605.10040#pg2>. By integrating pseudoforgery feature simulation with evidential uncertainty modeling, FADNet achieves superior generalization capabilities against previously unseen forgery techniques<ref:2605.10040#pg8>. The uncertainty-driven decision mechanism provides enhanced reliability and interpretability for real-world digital forensic applications<ref:2605.10040#pg8>. The synergy between uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. This approach demonstrates superior stability in calibration performance compared to existing baselines<ref:2605.10040#pg8>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>. This synergy allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>. This synergy between uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>. This synergy between uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>. This synergy between uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>. This synergy between uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>. This synergy between uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>. This synergy between uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>. This synergy between uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>. This synergy between uncertainty regularization and feature compactness allows FADNet to effectively characterize the intrinsic distribution of authentic facial features<ref:2605.10040#pg10>. The experiments reveal that while uncertainty regularization is indispensable for enabling facial forensics, feature compactness further refines and optimizes detection precision<ref:2605.10040#pg10>.

Improvements for AI systems

Improved AI system can perform universal face forgery detection by framing it as a one-class classification (OCC) task, as stated in Section I: we reformulate face forgery detection as a one-class classification (OCC) task. This allows the model to be trained exclusively on authentic facial data to capture intrinsic representations.
Improved AI system can enhance generalization against novel forgery techniques by incorporating a plug-and-play pseudo-forgery image generator (PFIG) during training, which tighten[s] decision boundaries around authentic data. This module applies four types of perturbations—including Spectral Disruptions, Local Blending, and Checkerboard Artifacts—to simulate forgery artifacts.
Improved AI system can provide calibrated uncertainty scores for its predictions by incorporating Evidential Deep Learning (EDL), which models uncertainty via the Dirichlet distribution. This allows the model to assign higher uncertainty to out-of-distribution (OOD) fake images, facilitating a natural separation between real and forged content.
Improved AI system can maintain high confidence calibration in forensic applications by using an uncertainty-based thresholding strategy during inference, where authentic face images yield low uncertainty values and others generate high values. This is supported by the finding that the model achieves optimal accuracy at a threshold of 0.691: The model achieves optimal accuracy and F1-score at a threshold of 0.691.
Improved AI system can ensure feature separability by incorporating a feature compactness loss, defined as Lcompact = (f(I(i)r) - ¯fr) 2 to minimize intra-class variance for real faces. This constraint helps in enhancing their separability from forgeries, as demonstrated by the ablation study showing that omitting this loss leads to a decrease in accuracy to 76.% and an F1-score of 86.87%.

Sources

Deepfake Generation and Detection: A Benchmark and Survey