EMSFD: Evidential Synthetic Face Detection

• EMSFD is an uncertainty-aware framework for synthetic-face detection that replaces Softmax with evidential deep learning to enhance calibration.
• It performs a two-stage process by first classifying images as real or synthetic, then attributing synthetic images to either GAN or diffusion sources using frequency-domain analysis.
• EMSFD integrates uncertainty-driven active learning, contrastive representation, and a robust ViT-B/16 backbone to achieve state-of-the-art accuracy and labeling efficiency.

Searching arXiv for the exact topic and nearby variants to ground the article in current papers. EMSFD denotes Evidence-based Decision Modeling for Synthetic Face Detection with uncertainty-driven active learning, an uncertainty-aware framework for synthetic-face forensics that replaces standard Softmax decision modeling with an evidential deep learning formulation and then uses the resulting uncertainty estimates to drive active learning (Jiang et al., 11 May 2026). The method targets two coupled requirements that are central in contemporary synthetic-face detection: reliable inference under out-of-distribution generator shift, and reduced annotation cost when new generators and artifacts emerge rapidly. In its reported formulation, EMSFD performs a two-stage synthetic face analysis task: Stage 1 classifies an input face as Real or Synthetic in the spatial domain, and Stage 2, conditioned on a synthetic decision, classifies the image as originating from a GAN or a diffusion model (DM) in the frequency domain (Jiang et al., 11 May 2026).

1. Nomenclature and scope

The acronym EMSFD refers specifically to synthetic face detection, not to multilingual sentence representation learning, emergency medical service datasets, EMS dialogue generation, or Earth-system forecasting. Contemporary arXiv papers use nearby acronyms for unrelated topics, including EMS for massively multilingual sentence embedding learning (Mao et al., 2022), EgoEMS for an egocentric EMS dataset (Weerasinghe et al., 13 Nov 2025), EMSDialog for synthetic EMS conversations (Ge et al., 8 Apr 2026), and ESFM for an Earth System Foundation Model (Ozdemir et al., 20 Apr 2026). In this usage, EMSFD is a computer-vision and machine-learning method for face-forensics rather than a dataset, dialogue framework, or foundation model outside the image-forensics domain.

Within synthetic-face analysis, EMSFD addresses both binary authenticity detection and generator-family attribution. The problem setting is explicitly organized as a cascade. First, the model decides whether a face is real or synthetic. Second, if synthetic, a separate classifier identifies whether the source belongs to the GAN family or the diffusion-model family. This decomposition is motivated by the observation that different generative paradigms leave distinct traces, and that frequency-domain discrepancies are especially useful for source attribution (Jiang et al., 11 May 2026).

2. Problem formulation and motivation

EMSFD is motivated by two limitations that the paper assigns to conventional synthetic-face detectors. The first is overconfidence induced by Softmax classifiers. For an input sample $i$, a conventional detector predicts class probabilities from logits $z_{ij}$ by

$$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$

The paper argues that this normalization can amplify small logit differences into sharply peaked probabilities, forcing the model to allocate mass to one of the known classes even for unknown out-of-distribution images. In the reported analysis, this leads to unreliable confidence, poor calibration, and the inability to express ignorance when confronted with unseen generators (Jiang et al., 11 May 2026).

The second limitation is dependence on large labeled datasets. The paper treats this as a practical barrier because new image generators appear quickly, synthetic distributions shift, and annotation of diverse synthetic data is costly. EMSFD therefore couples uncertainty-aware prediction with an active learning loop that prioritizes high-uncertainty unlabeled samples for annotation. The intended outcome is simultaneous improvement in reliability, generalizability, and label efficiency (Jiang et al., 11 May 2026).

The task design also exploits domain structure. Stage 1 operates on RGB face images in the spatial domain, while Stage 2 operates on a frequency-domain representation to distinguish GANs from diffusion models. This two-stage arrangement is presented as a way to combine authenticity detection with source attribution rather than collapsing both problems into a single flat classifier.

3. Evidential decision modeling

The conceptual core of EMSFD is the replacement of point-probability classification with evidence-based decision modeling. For an input $x_i$ and $K$ classes, the network produces a nonnegative evidence vector

$\mathbf{e}_i=[e_{i1},e_{i2},\ldots,e_{iK}]^T,$

and converts evidence into Dirichlet parameters by

$\alpha_{ik}=e_{ik}+1.$

This yields a Dirichlet distribution over class probabilities,

$p_i \sim \mathrm{Dir}(\boldsymbol{\alpha}_i),$

with density

$$D(p_1,\ldots,p_K\mid \alpha)=\frac{1}{B(\boldsymbol{\alpha})}\prod_k p_k^{\alpha_k-1}, \qquad B(\boldsymbol{\alpha})=\frac{\prod_k \Gamma(\alpha_k)}{\Gamma\left(\sum_k \alpha_k\right)}.$$

The Dirichlet strength is

$S=\sum_k \alpha_k,$

and the paper uses

$z_{ij}$0

In this parameterization, large total evidence corresponds to concentrated predictions and low uncertainty, whereas low total evidence corresponds to diffuse predictions and high uncertainty (Jiang et al., 11 May 2026).

The reported implementation requires the evidence to be nonnegative. The paper considers ReLU, Softplus, and Exponential activations, and selects Exponential in the final model because it performs best in ablation. This evidential representation is then optimized with a supervised Dirichlet-based loss,

$z_{ij}$1

where $z_{ij}$2 is batch size, $z_{ij}$3 is a sample weight for class imbalance, $z_{ij}$4 is the one-hot target, $z_{ij}$5, and $z_{ij}$6 is the digamma function (Jiang et al., 11 May 2026).

To improve feature geometry, EMSFD adds a supervised contrastive term,

$z_{ij}$7

where $z_{ij}$8, $z_{ij}$9, and $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$0 are $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$1-normalized embeddings, $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$2, and $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$3. The total objective combines the two losses with equal weights: $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$4 A plausible implication is that EMSFD treats uncertainty not as a post-hoc calibration layer but as a first-class training signal that structures both prediction and representation learning.

4. Architecture and uncertainty-driven active learning

The main architectural component is the Face Evidence Extraction (FEE) module, which consists of a frozen backbone, a linear layer for projection or dimensionality reduction, and an evidence layer producing nonnegative evidence. The primary backbone is ViT-B/16. The paper also compares ResNet-101, EfficientNet-B3, and DenseNet-121, but ViT-B/16 is retained as the final design because it gives the best reported accuracy and calibration (Jiang et al., 11 May 2026).

In Stage 1, FEE receives an RGB face image and predicts Real versus Synthetic. In Stage 2, the input image is transformed with a 2D FFT independently on each RGB channel. The complex-valued FFT output is decomposed into log-magnitude $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$5 and phase $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$6, producing a 6-channel frequency representation formed by concatenating the 3-channel log-magnitude and 3-channel phase maps. The ViT patch-embedding layer is modified to accept these 6 input channels (Jiang et al., 11 May 2026).

The active learning mechanism uses evidential uncertainty directly. Let $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$7 denote the labeled set and $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$8 the unlabeled pool. For each unlabeled sample $$p_{ij}=\frac{\exp(z_{ij})}{\sum_{k=1}^{K}\exp(z_{ik})}.$$9, EMSFD computes

$x_i$0

ranks the pool by uncertainty, and selects the top-$x_i$1 samples,

$x_i$2

After annotation, the dataset partitions are updated by

$x_i$3

The reported protocol uses half of the training set as labeled data and half as unlabeled at initialization; in each round the model trains for 10 epochs, queries the top 20% most uncertain unlabeled samples, and repeats this process for 5 rounds in both stages (Jiang et al., 11 May 2026).

Implementation details reported for the final setup include AdamW, initial learning rate $x_i$4, weight decay $x_i$5, Cosine Annealing with period 100 steps, gradient clipping with max norm 1.0, batch size 16, and training on a single NVIDIA RTX 5060Ti GPU. The augmentation pipeline uses Random Resized Crop, Random Horizontal Flip, Random Rotation, Random Affine, and Color Jitter (Jiang et al., 11 May 2026).

5. Experimental protocol and empirical results

The reported experiments use Asian faces from FFHQ as real images and the ASFD dataset for synthetic faces. To evaluate cross-generator generalization, the authors train with StyleGAN and ADM, and test on unseen out-of-distribution generators StyleGAN2, VQGAN, IDDPM, and LDM. They also evaluate an additional cross-paradigm setting using FLUX.2, which is based on Flow Matching rather than GANs or diffusion models; because it does not fit the GAN-versus-DM taxonomy, this experiment is limited to Stage 1 binary detection (Jiang et al., 11 May 2026).

For Stage 1, the training set contains 6,000 images: 2,000 real, 2,000 GAN-generated, and 2,000 DM-generated. During training, GAN and DM images are merged into the synthetic class, so Stage 1 is trained on 2,000 real and 4,000 synthetic images. For Stage 2, the training set contains 4,000 synthetic images: 2,000 GAN-generated and 2,000 DM-generated. Under active learning, the reported splits are 3,000 labeled + 3,000 unlabeled for Stage 1 and 2,000 labeled + 2,000 unlabeled for Stage 2. Evaluation metrics are Accuracy (Acc), Average Precision (AP), and Expected Calibration Error (ECE) (Jiang et al., 11 May 2026).

The central comparison reports the following average results across evaluated generators: CNNSpot at 60.5% Acc, 84.2% AP; LGrad at 56.5% Acc, 48.2% AP; DIRE at 69.1% Acc, 84.8% AP; DeepFeatureX at 66.9% Acc, 82.4% AP; Cutting-Edge at 59.9% Acc, 81.5% AP; MDL at 81.6% Acc, 89.5% AP; and EMSFD at 96.4% Acc, 96.2% AP (Jiang et al., 11 May 2026). Relative to the strongest baseline in average accuracy, MDL at 81.6%, EMSFD improves to 96.4%, corresponding to 14.8 percentage points, effectively about 15%.

The OOD results are particularly pronounced. On VQGAN, CNNSpot achieves 1.4% Acc, DIRE 3.6%, Cutting-Edge 4.4%, MDL 25.5%, while EMSFD reaches 86.1%. On LDM, EMSFD reaches 99.3% Acc / 97.9% AP. Calibration results on OOD data show ECE values of 0.6179 for CNNSpot, 0.5803 for DIRE, and 0.0516 for EMSFD, which supports the paper’s claim that the model is not only more accurate but substantially better calibrated (Jiang et al., 11 May 2026).

The ablation studies isolate the evidential design choices. For evidence activation, Softplus gives 87.5% average Acc, 97.0% average AP, ReLU gives 89.9% average Acc, 96.8% average AP, and Exponential gives 96.4% average Acc, 96.2% average AP. For backbones, ViT-B/16 yields 96.4% Acc, 96.2% AP, 0.052 ECE, compared with ResNet-101 at 87.1% Acc, 89.0% AP, 0.297 ECE, EfficientNet-B3 at 89.6% Acc, 92.9% AP, 0.154 ECE, and DenseNet-121 at 91.2% Acc, 92.9% AP, 0.226 ECE. A direct Softmax versus evidential comparison on the same ViT backbone reports 89.1% average Acc, 93.2% average AP, 0.203 ECE for Softmax and 96.4% average Acc, 96.2% average AP, 0.052 ECE for the Exponential evidential head (Jiang et al., 11 May 2026).

Under active learning, the paper compares CNNSpot + AL at 88.0% average Acc, 91.6% average AP with EMSFD + uncertainty-driven AL at 96.4% average Acc, 96.2% average AP. This suggests that the informativeness of the uncertainty estimate, rather than the presence of active learning alone, is a central part of the reported gain.

6. Interpretability, limitations, and practical significance

EMSFD presents interpretability primarily through uncertainty-aware evidence strength and feature-space organization. The paper reports t-SNE visualizations for both stages in which class clusters are compact and well separated, with low uncertainty near cluster centers and high uncertainty near class boundaries. It also notes that average uncertainty is higher in Stage 2 than in Stage 1, which is consistent with the greater difficulty of source attribution relative to binary authenticity detection (Jiang et al., 11 May 2026).

The method’s reported strengths are explicit: explicit uncertainty modeling through a Dirichlet evidence framework, reduced overconfidence relative to Softmax detectors, better calibration, stronger OOD generalization across unseen generators, improved labeling efficiency through uncertainty-driven active learning, interpretability, and a two-stage design that exploits both spatial and frequency cues. The paper characterizes these properties as particularly relevant for real-world synthetic-face detection, where new generators appear frequently, annotation is expensive, and false high-confidence decisions are dangerous (Jiang et al., 11 May 2026).

The reported limitations are equally specific. The system is cascaded, so Stage 2 depends on Stage 1, even though the authors argue that error propagation is negligible because Stage 1 reaches near-saturated accuracy. The active learning strategy is deliberately simple, relying only on uncertainty ranking rather than diversity or representativeness criteria. Evaluation is conducted on a curated set of generators plus one additional unseen paradigm, so broader open-world testing remains desirable. The uncertainty formulation is effective but does not decompose uncertainty into aleatoric and epistemic components. The active learning setup still requires a substantial initial labeled set, since half of the training data is labeled at the start. Finally, the FLUX.2 results—64.9% Acc, 67.4% AP, 0.346 ECE for EMSFD, versus 54.1% Acc, 18.9% AP, 0.459 ECE for CNNSpot and 0.0% Acc, 0.0% AP, 1.000 ECE for DIRE—show that genuinely novel paradigms remain difficult, even when EMSFD generalizes better than the baselines (Jiang et al., 11 May 2026).

The code release is reported at https://github.com/hzx111621/EMSFD. Taken as a whole, EMSFD is best understood as a synthetic-face detector that integrates evidence-based uncertainty estimation, contrastive representation learning, frequency-aware attribution, and uncertainty-driven sample acquisition into a single framework. Its defining claim is not merely improved mean accuracy, but a joint improvement in accuracy, calibration, interpretability, and label efficiency under generator shift.