Back to paper library
arXiv 2024CVPR Workshops 2024Face anti-spoofingPaper notes

FAS-Challenge

Arsenal

Original paper

Unified Physical-Digital Attack Detection Challenge

Paper authors

Haocheng Yuan · Ajian Liu · Junze Zheng …

Show all 9 authors

Haocheng Yuan · Ajian Liu · Junze Zheng · Jun Wan · Jiankang Deng · Sergio Escalera · Hugo Jair Escalante · Isabelle Guyon · Zhen Lei

Publication
CVPR Workshops 2024Conference
First submitted to arXiv
arXiv:2404.06211
Cite this paper · BibTeX

The citation identifies the linked source version. It cites the original paper and its authors.

@misc{faceantispoofingfaschallenge2024,
  title = {{Unified Physical-Digital Attack Detection Challenge}},
  author = {Haocheng Yuan and Ajian Liu and Junze Zheng and Jun Wan and Jiankang Deng and Sergio Escalera and Hugo Jair Escalante and Isabelle Guyon and Zhen Lei},
  year = {2024},
  eprint = {2404.06211},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.06211}
}
On this page

This is the FAS competition held at CVPR2024, officially titled:

5th Chalearn Face Anti-spoofing Workshop and Challenge@CVPR2024

It’s a lively event that everyone working in the FAS field can check out.

Problem Definition

The essence of the Face Anti-Spoofing (FAS) task is to mine the "syntax" of liveness from images.

However, in reality, attack methods have evolved into two parallel technical branches:

  • Physical Attacks (PA): The impersonator presents the face through physical media such as paper, screens, or silicone masks. These interferences mostly occur at the sensor level and have direct interaction with the real world.
  • Digital Attacks (DA): Such as Deepfake, face swapping, or adversarial examples, which manipulate from the data generation source or feature level. Although visually realistic, their essence is synthetic.

Traditional methods mostly model one of these categories, thus model performance is limited to their training distribution and cannot generalize broadly.

Although both physical and digital attacks belong to the “fake” class in final classification, their image statistical features and variation directions are highly heterogeneous. This causes intra-class feature distances far larger than expected, leading to a generalization bottleneck.

Currently, the core reasons why it is difficult to build a "unified model" are twofold:

  1. Lack of large-scale unified datasets: Past works typically splice separately collected PA and DA datasets, without covering complete attack types for the same ID.
  2. Absence of public evaluation benchmarks: Physical and digital attacks use different metrics and protocols, making consistent comparison across domains impossible.

This is the background for initiating the Unified Physical-Digital Attack Detection Challenge. Through new datasets, standard protocols, and open competitions, it attempts to define a new problem setting:

Can a single model handle two heterogeneous spoofing types simultaneously while maintaining discriminative power in unseen domains?

Unified Dataset: UniAttackData

At the time of this paper, UniAttackData was described as the largest public dataset unifying physical and digital attacks. It covers 1,800 subjects and 28,706 facial video clips in three sample categories; the figure above shows how the challenge splits them by protocol:

  • Live: 1,800 genuine video clips
  • Physical Attacks (PA): 5,400 clips from print and screen replay attacks.
  • Digital Attacks (DA): 21,506 clips, including Deepfake, Face Swap, adversarial examples, etc.

The dataset’s key feature is that each ID covers the full set of attack types. This reduces the risk of confounding identity with attack type; the live and spoof counts in each protocol remain imbalanced.

The challenge organizers cropped face regions from the videos and selected one frame per video. They then renumbered filenames to hide clues, such as attack type, that the original names might reveal.

Evaluation Protocols and Generalization Design

The challenge includes two main protocols simulating different real-world deployment scenarios:

  • Protocol 1: Unified Attack Detection Simulates scenarios where the model must simultaneously recognize both PA and DA, testing the model’s integration and classification capability for mixed attack types.

  • Protocol 2: Generalization to Unseen Attacks Zero-shot testing for “unseen attack types,” further divided into:

    • Protocol 2.1: Unseen Physical Attacks — no PA in training; PA in the test set.
    • Protocol 2.2: Unseen Digital Attacks — no DA in training; DA in the test set.

This protocol withholds the broad physical or digital attack category until testing. It does not test every fine-grained attack subtype in a separate leave-one-out run.


Competition Process and Rules

The overall competition runs in two phases on the CodaLab platform:

  • Development Phase (2/1–2/22) Provides labeled training data and unlabeled development set; participants can repeatedly submit predictions to the leaderboard for iterative model tuning.

  • Final Phase (2/23–3/3) Reveals development set labels and releases an unlabeled test set. Teams must submit final predictions without access to true test labels. The last submission counts as the official score, and teams must publicly release their code and fact sheets to qualify for awards.


Evaluation Metrics

The challenge uses the ISO/IEC 30107-3 international standard metrics to quantify model spoof detection ability, including:

  • APCER (Attack Presentation Classification Error Rate)
  • BPCER (Bona Fide Presentation Classification Error Rate)
  • ACER (Average Classification Error Rate)
  • AUC (Area Under the ROC Curve)

The main ranking metric is ACER, with AUC as a secondary indicator. To maintain evaluation consistency, the final ACER is computed using a threshold calibrated by the Equal Error Rate (EER) on the development set.


Arsenal

Next is a technical overview of the top 13 participating teams. Let’s take a closer look at what they actually did.

1. MTFace

In this challenge spanning both digital and physical attacks, the MTFace team proposed the ultimate winning solution.

The MTFace architecture can be briefly described as:

Optimized Data Augmentation for Comprehensive Face Attack Detection Across Physical and Digital Domains

The name sounds long but captures the core issue: how to expose the model during training to "sufficiently diverse and realistic" spoof appearances.

The core of MTFace lies in the joint design of data augmentation and loss balancing.

Its data preprocessing steps are as follows:

  1. Face detection and cropping: Detect faces in all images, with an additional 20-pixel margin to retain peripheral features.
  2. Face mask extraction: Pre-extract masks of live data faces for subsequent augmentation use.

Then, for the generalization requirements of different protocols, two "semantically aligned augmentation strategies" were designed:

  • Moire simulation augmentation (for Protocols 1 and 2.1) Screen replay often causes moire patterns, which become important visual cues for attack recognition. MTFace simulates this interference by injecting moire patterns into original images, enabling the model to anticipate this physical phenomenon’s potential impact.

  • Self-fusion augmentation (for Protocols 1 and 2.2) Inspired by previous literature, MTFace uses live data as a base and injects surface features of digital attacks. Through color distortion, spatial deformation, and mask folding, it creates "hybrid samples of digital spoofs," improving the model’s discrimination against Deepfake-type attacks.

The ratio between genuine and attack samples varies greatly across protocols.

Without adjustment, the model tends to bias toward the more numerous class, losing the ability to recognize rare types.

To address this, MTFace adjusts cross-entropy weight configurations per protocol:

  • Protocol 1: live : fake = 1 : 1 Maintain balanced loss to ensure unbiased learning.
  • Protocol 2.1 (Unseen PA): live : fake = 5 : 1 Increase the loss weight for live faces before testing on unseen physical attacks.
  • Protocol 2.2 (Unseen DA): live : fake = 2 : 1 Adjust the influence of live and fake samples to suit this protocol’s sample ratio.

Together with data augmentation, these weights are part of MTFace’s training strategy across the three protocols. The paper does not isolate the effect of each loss ratio in a separate ablation.

MTFace ultimately adopts ResNet-50 as backbone, loading ImageNet pretrained weights.

No ViT, no custom large model—just strategy and data, clinching first place in this competition.

2. SeaRecluse

SeaRecluse chose to build defenses through data boundaries and proportion configurations.

Their solution is named:

Cross-domain Face Anti-spoofing in Unified Physical-Digital Attack Dataset

This architecture does not emphasize extreme transformations or style transfer. Instead, it focuses on data usage ratios and cropping strategies to optimize model stability and generalization in cross-domain tasks, closer to real deployment conditions.

SeaRecluse uses SCRFD to detect and crop faces from uncropped images in the training set.

Unlike other teams, they distinguish and supplement between loose and tight crops as one form of data augmentation.

Additionally, for different task protocols, data split ratios and augmentation strategies differ completely:

  • Protocol 1: No additional augmentation Simulate baseline recognition performance, training on 80% training data mixed with validation set.

  • Protocol 2.1 (Unseen PA): Downsample and pad live data edges, expanding real face data to 3× the original to balance real and fake samples.

  • Protocol 2.2 (Unseen DA): To enhance model perception of fake samples, downsample fake faces by 4× and 8×, totaling 7× the original amount.

The team also corrected some images with abnormal aspect ratios to restore reasonable proportions, preventing the model from learning erroneous biases due to visual distortion.

For image augmentation, all tasks applied standard operations (flipping, random cropping, etc.), with Protocol 2.1 additionally using Gaussian Blur to simulate photographic blur and distant-view defocus.

The backbone network is ConvNeXt V2; the team says challenge requirements and local training resources informed that choice.

To further strengthen generalization, the team applied two training techniques:

  • Image CutMix: Mixes two images and labels to improve adaptation to visual boundaries and spatial variations.
  • Label Smoothing: Converts hard labels to soft labels to reduce overfitting risk.

The team uses these techniques to augment training; their separate effects on class imbalance require ablation.

SeaRecluse’s approach is not drastic but rather wall-building—starting from task-specific data ratios and cropping scales to confine spoofing within the model’s field of vision.

It is a patient way to handle data splits and cropping scales.

3. duileduile

The duileduile team designed a modular two-stage learning pipeline that combats the fluid boundary between real and spoof through abstract visual representations. The core of their architecture is:

Swin Transformer + Masked Image Modeling + Physical and Digital Attack Corresponding Augmentation

The backbone model is Swin-Large Transformer, extracting a 1536-dimensional feature vector, with strong regional awareness and hierarchical abstraction capabilities.

In the pretraining stage, duileduile employs the simMIM (Simple Masked Image Modeling) strategy, which partitions the image into non-overlapping patches and randomly masks some parts, forcing the model to learn “reconstructing the whole from incomplete inputs.”

This self-supervised pretraining uses unlabeled images from the training and development sets to learn representations from masked inputs. The paper does not separately measure its contribution to occlusion attacks or Protocol 2.

After training the visual syntax, the process moves to the fine-tuning stage. At this point, the strategy shifts from large-scale data stacking to precise pattern fitting per attack type, with corresponding augmentation procedures designed as follows:

  • Gaussian Noise: Simulates pixel-level noise and compression artifacts caused by digital attacks
  • ColorJitter + Moire Pattern + Gamma Correction: Reconstructs light and shadow variations and display biases of physical attacks

These augmentations are not uniformly applied to all data but probabilistically applied according to different training samples and protocol tasks, exposing the model to varied disturbances and spoofs each time it learns.

Compared with the first two teams’ focus on augmentation and data ratios, duileduile keeps its two-stage backbone and training configuration the same across protocols. That is one implementation choice in this challenge.

4. BSP-Idiap

The BSP-Idiap team took a different path: returning to the signal’s intrinsic texture rhythm by reconstructing attack signals in the frequency domain.

Their method is called: DBPixBiS (Dual-Branch Pixel-wise Binary Supervision)

Building upon their prior design philosophy, they expanded it into a dual-branch architecture.

DBPixBiS employs a dual-branch neural network structure:

  1. RGB branch: Uses Central Difference Convolution (CDC) instead of traditional convolution, emphasizing local texture changes and enhancing the model’s perception of abnormal edges and subtle variation areas.
  2. Fourier branch: Applies Fourier transform to input images and feeds them into a separate feature path, capturing spoofing textures in the frequency domain, such as repetitiveness, interference, and compression residues.

This design enables the model to simultaneously perceive “visual anomalies” in the image and “signal-level spoofing” within the data.

To mitigate overfitting and the subtlety of adversarial examples, BSP-Idiap adopts highly targeted training designs:

  • Pixel-wise Binary Supervision: Supervises binary classification at each pixel on the feature map instead of the entire image, improving recognition of localized spoofing.
  • Attentional Angular Margin Loss: Adds an angular margin penalty during training, guiding the model to distinctly separate live and spoof feature vectors, reinforcing stable class boundaries.

During testing, the model performs global average pooling on the feature map and uses the mean activation as the final spoof probability score.

5. VAI-Face

In this unified recognition battle, some teams build mechanisms, some restore syntax, and the VAI-Face team chose a third way: create anomalies to strengthen recognition.

Their solution centers on the Dinov2 Vision Transformer (ViT-Large) architecture, aiming to expose flaws in distorted faces through visual attention.

Their key data augmentation strategy is the deliberate asymmetric treatment of live and fake images.

They treat the two as semantically heterogeneous sources and thus apply drastically different augmentation pipelines:

  • Live images: Only undergo RandomResizedCrop and HorizontalFlip, preserving natural distribution and geometric stability.
  • Fake images: Subject to heavy asymmetric perturbations such as blur, distortion, and custom cutout occlusions, simulating various unnatural traces and structural breakages common in spoof images.

This strategy trains the ViT model to identify abnormal geometry and texture.

Beyond data strategies, VAI-Face exhibits high engineering sensitivity in training configuration:

  • OneCycleLR: Precisely controls learning rate ramp-up and decay to improve convergence efficiency and generalization.
  • Label Smoothing: Prevents overconfidence on specific patterns, reducing overfitting.
  • Mixup Augmentation: Combines two images and labels to improve robustness at sample space boundaries.
  • Optimizer: Uses ADAN, a novel optimizer combining adaptive gradients with momentum, providing more stable gradient dynamics in this challenge.

ViT-Large inherently possesses strong regional relational modeling ability. When processing spoof images, its global attention mechanism captures subtle inconsistencies, amplifying flaws hard to conceal. They do not add extra branches or generative modules but maximize the backbone’s recognition potential through finely tuned data perturbations and learning schedules, squeezing out generalization from a simple architecture.

6. L&L&W

While most models choose holistic input and full-image classification, the L&L&W team took the opposite approach, starting from local details: slicing a face into countless fragments and piecing together subtle clues of authenticity or spoof.

Their core strategy revolves around:

Patch-based Feature Learning + Frequency-guided Sampling + Local Attention Guidance

The process begins with image patch extraction, splitting each input image into multiple small regions for independent feature learning. On this foundation, they introduce the Centre Difference Attention (CDA) mechanism.

CDA is an attention method enhancing subtle texture differences, focusing on regions that "should be consistent but show minor variations," such as edge halos, misaligned reassembly, and low-frequency fusion failures.

Beyond spatial details, L&L&W also leverage hidden frequency-domain clues. They designed the High-Frequency Wavelet Sampler (HFWS) module to concentrate on high-frequency bands in images, aiming to detect compression artifacts, fusion distortions, and unnatural texture discontinuities left during forgery.

This feature strategy brings local attention and high-frequency components into the recognition process. The paper does not separately measure how much each of these two modules contributes.

During validation and testing, the team generates 36 cropped versions per image and averages their prediction scores into the final score.

7. SARM

The SARM team’s design focuses not on complex structures or data augmentation but on a repeated “training the trainer” process.

Their method is called:

Multi-Attention Training (MAT) + Label Flip Augmentation

This is a staged visual refinement approach: first, train a model to understand the task, then use this model to finely tune and adapt to different attack types.

SARM splits the process into two phases:

  1. Phase One: Train dedicated pre-detectors for each protocol (P1, P2.1, P2.2) using Supervised Contrastive Learning to maximize semantic distance between live and fake representations.

  2. Phase Two: Fine-tune the actual anti-spoofing model starting from phase one’s representations as initial weights.

This design strengthens “expected object understanding” per task, reducing misleading effects common in direct end-to-end training, especially effective for cross-domain generalization tasks (P2.1 and P2.2).

In data augmentation, SARM innovates by transforming live faces into fake ones and labeling them as fake.

This is not ordinary synthetic spoofing but a weak spoof simulation using OpenCV style transformations.

For training data in P2.1 and P2.2, they use OpenCV-based image style transformations. The paper does not list separate hue, lighting, gamma, or filter settings.

These processed live images are labeled as spoof, creating label-flip augmentation training samples. The authors argue that this can narrow the train–test domain gap, but they do not define the resulting images as a separate attack category.

The optimizer is Adam, combined with cross-entropy and supervised contrastive loss. The paper says P1 uses the ordinary training strategy and P2.1/P2.2 additionally use label-flip augmentation; it does not provide convergence curves for these stages.

This method mainly relies on "prior task intention understanding" to drive model convergence.

8. M2-Purdue

The M2-Purdue team offers a different strategy.

They abandoned complex architecture design and heavy data augmentation, opting to use CLIP’s semantic representations combined with an “extreme risk–oriented” loss, named:

Robust Face Attack Detection with CLIP + MLP + CVAR–AUC Loss Fusion

The pipeline starts with standard image preprocessing, resizing all inputs to 224×224 to ensure scale consistency.

Next, they extract semantic features via CLIP’s image encoder, converting visual information into deep embedding representations.

The key here is not creating new features but “leveraging a model pretrained on massive data” to obtain CLIP visual features. They then attach a three-layer MLP classifier as the task-specific decision module. The paper says the MLP is fine-tuned; the full update scope and mobile deployment cost require further verification.

The most distinctive aspect is the dual loss design, integrating two supervisory signals:

  • CVAR (Conditional Value at Risk) Loss: Common in financial risk control, focusing on tail risk—the hardest-to-classify, highest-risk sample regions.
  • AUC (Area Under Curve) Loss: Emphasizes overall discriminative ability, optimizing the model’s perception of correct ranking.

The overall loss is defined as:

L=λ⋅LCVAR+(1−λ)⋅LAUC\mathcal{L} = \lambda \cdot \mathcal{L}_{\mathrm{CVAR}} + (1 - \lambda) \cdot \mathcal{L}_{\mathrm{AUC}}

where parameter λ\lambda balances “extreme risk sensitivity” and “overall generalization.”

The underlying logic is: not striving for all-round correctness, but minimizing errors in the “most error-prone zones.”

The paper’s prose says training uses Adam, while the architecture figure labels a SAM optimizer; the challenge report does not resolve which was actually used. This method combines CLIP features with tail-risk and ranking losses, but the paper does not provide independent convergence-stability or deployment measurements.

9. Cloud Recesses

If the essence of anti-spoofing is finding flaws on the face, then the Cloud Recesses team’s strategy provides a creative twist: masking the most recognizable parts, forcing the model to understand authenticity from subtler signals.

Their method is called:

Random Masking for Face Anti-Spoofing Detection

This is a data-level adversarial training approach that directly obscures key facial regions, forcing the model to learn liveness recognition even without clear views of eyes and mouth.

The overall process breaks down into three steps:

  1. Face Detection: Use RetinaFace to crop faces from original images, standardizing them to 256×256 size.
  2. Keypoint Annotation: Use dlib to detect 68 facial landmarks, precisely outlining eyes, nose tip, lips, etc.
  3. Random Masking: For each training sample, randomly mask three to five key regions, depriving the model of shortcut cues relying on facial features.

This design covers local regions such as the eyes, nose, and mouth, reducing reliance on particular facial-part cues. The paper does not verify which alternative textures the model learns, nor does this method use video motion blur as a feature.

The backbone is EfficientNet, balancing accuracy and efficiency, suitable for lightweight recognition on masked images.

Training uses random masking as data augmentation so the model must classify faces with some local information missing. The paper does not separately report an ablation of the masking strategy.

Cloud Recesses’ solution involves no extra modules, no dual branches, no feature fusion — just removing key visual cues from the face to test if the model can recognize truth from darkness.

10. Image Lab

In anti-spoofing detection tasks, we usually want the model to "see clearly," but the Image Lab team chose to have the model "see many times." Their architecture is called:

Multiattention-Net: A Deep Visual Recognition Network Composed of Multiple Attention Layers

This design systematically integrates multi-stage spatial information and multiple attentions, extracting possible anomaly cues from different dimensions of the image.

The network starts with a 7×7 convolutional layer capturing local texture details, followed by ten modified squeezed residual blocks. The diagram shows max pooling between groups of residual blocks, progressively abstracting the input into increasingly global spatial features.

Throughout this process, spatial information is continuously extracted at each layer, finally followed by a Dual Attention Block that strengthens the recognition weights of key regions. This enables the model to discover spoofing traces both in fine details and structure.

Finally, Global Average Pooling (GAP) reduces the attention branch outputs, which are concatenated before entering a fully connected layer for classification.

During training, Image Lab uses Binary Focal Cross Entropy Loss, a loss function designed to penalize minority classes and high-confidence errors more strongly, with the formula:

L(y,y^)=−α⋅(1−y^)γ⋅log⁡(y^)−(1−α)⋅y^γ⋅log⁡(1−y^)\mathcal{L}(y, \hat{y}) = -\alpha \cdot (1 - \hat{y})^\gamma \cdot \log(\hat{y}) - (1 - \alpha) \cdot \hat{y}^\gamma \cdot \log(1 - \hat{y})

where:

  • y^\hat{y} is the predicted probability, yy is the ground truth;
  • α=0.25\alpha = 0.25 (when class balancing is enabled);
  • γ=3\gamma = 3 (to emphasize hard samples).

The paper prints two probability terms in Equation (2), but does not switch between positive and negative classes according to the true label yy. That printed expression alone is insufficient to reconstruct the actual binary focal loss. ImageLab’s final ACER is 34.9434%, and the paper does not separately measure convergence efficiency.

11. BOVIFOCR-UFPR

In this battle focused on "recognizing spoofs," BOVIFOCR-UFPR is among the few teams extending analysis beyond the 2D plane. Rather than only analyzing image colors and textures, they attempt to reconstruct the 3D authenticity of the entire face, thereby detecting spatial inconsistencies that spoofs cannot reproduce.

Their core solution is:

3D Reconstruction + ArcFace + Chamfer Loss

Inspired by 3DPC-Net, the framework centers on an Encoder-Decoder architecture:

  1. Preprocessing: Detects, aligns, and crops the face, and obtains reference 3D point clouds for training through an existing reconstruction method.
  2. Encoder: Employs ResNet-50 backbone to extract high-level features.
  3. Decoder: Transforms features into corresponding 3D point cloud representations, simulating the face's geometric shape in space as a basis for subsequent discrimination.

With this structure, the model learns not only "how similar it looks" but "how plausible it is," verifying spatial consistency.

During training, the team combines two feature-guided losses:

  • ArcFace Loss: Adds an angular-margin constraint to classification features. The figure’s output is live/fake classification, so it does not show that the model identifies people.
  • Chamfer Distance Loss: Measures spatial distance between predicted point clouds and pre-reconstructed reference point clouds to constrain geometry.

The model must classify live/fake and simultaneously output a 3D geometry consistent with reality.

Among the thirteen teams described here, this one uses 3D point-cloud reconstruction to assist live/fake classification. Its final ACER is 35.3620%.

It is a cross-dimensional recognition approach, relying on geometric plausibility rather than pixel details to build defenses.

12. Inria-CENATAV-Tec

The Inria-CENATAV-Tec team returned to a classic question:

Can stable anti-spoofing be achieved under limited computational resources?

Their answer is: MobileNetV3-spoof with hyperparameter tuning

This is a dynamic balance experiment between model complexity and recognition accuracy.

The pipeline reflects their systematic and conservative strategy:

  1. Facial landmark detection: Uses ResNet-50 for landmark detection.
  2. Alignment and fallback: If landmarks are detected, use InsightFace template for precise alignment; otherwise, keep original image resized.

This approach balances rigor (precise alignment when landmarks are available) and fault tolerance (does not discard data without landmarks), showing practical understanding of real-world scenarios.

Choosing MobileNetV3-large1.25 as backbone is key.

This architecture is optimized for edge AI and low-power devices, introducing SE attention blocks and h-swish nonlinear activation to balance parameters and recognition power.

With SGD optimizer + Multi-step Learning Rate scheduler, the model converges steadily in stages rather than through a one-shot global descent.

Besides basic augmentations (cropping, flipping), the team normalizes mean and standard deviation per protocol dataset to prevent feature shifts caused by domain heterogeneity.

This protocol-specific preprocessing is part of the team’s submitted method. The paper does not isolate its effect on generalization; the team’s final ACER is 37.3436%.

13. Vicognit

Among all participating teams, Vicognit is one of the few that explicitly focused on Vision Transformers (ViT).

Their solution is named:

FASTormer: Leveraging Vision Transformers for Face Anti-Spoofing

This design uses a Transformer as its backbone. The team aims to use its ability to encode relationships and capture sequential structure to learn cues for distinguishing live from spoof faces in spatial information.

The report describes Vicognit’s Transformer approach as distinguishing live and spoof faces without reducing the dimensionality of the input data. It does not provide an ablation of input resolution or compression.

Though not novel per se, this strategy is challenging in face anti-spoofing tasks.

Since differences between real and fake faces often lie not in geometry but in texture, material, and subtle perturbations, ViT’s global relational modeling fits well to learn these fine and irregular semantic gaps.

Their training strategy includes:

  • Carefully tuning learning rate and weight decay for stable, precise convergence;
  • The report does not detail any additional augmentation pipeline, so its contribution to the result cannot be assessed.

This approach uses a Transformer as the recognition backbone, but the final ACER is 52.1031% and AUC is 46.7025%; these results do not support a claim of strong generalization in this challenge.

Vicognit provides a comparison point for a pure Transformer architecture in this challenge.

Discussion

According to the official final statistics, the top five teams show several results worth noting:

  1. The top three teams clearly lead the others on ACER: the third-place team scored 5.5111%, versus 16.2263% for the fourth.
  2. The first-place team leads in ACER, AUC, and BPCER, but the best APCER was claimed by the fifth-place team, showing different models have selective advantages on different error types.
  3. The top five include teams from industry and research institutes; for example, fourth-place BSP-Idiap is from the Idiap Research Institute. The teams' backgrounds alone cannot explain the difference in scores.
  4. ACER varies widely across teams, from 2.3396% to 52.1031%. One competition is insufficient to identify any one architecture as a generally best solution.

Beyond ranking the teams, the challenge lets us compare different model and data strategies under the same protocols.

The organizers' post-competition summary suggests three directions worth testing further:

  • Path One: Examine the relationship between backbone capacity and results. Larger models may help, but the winner used ResNet-50, so size alone is not a sound selection rule.
  • Path Two: Improve data preprocessing and class balance. Several teams used cropping, augmentation, or loss weighting, but their contributions still need to be tested separately.
  • Path Three: Learn from images with local information missing. Masking and local-feature methods offer research directions; this report has no cross-method ablation that ranks them.

These are directions the organizers drew from the submissions.

Conclusion

Based on observations and organizers’ summaries of this competition, future UAD development still requires breakthroughs in three major directions:

  1. More comprehensive dataset construction: Although UniAttackData has laid the foundation for unified cross-attack datasets, improvements remain needed in attack type diversity, subject variability, and image quality. Especially with emerging adversarial attacks and style-transferred deepfakes, current sample sizes are insufficient for systematic generalization validation.

  2. Visual Language Model (VLM)-guided generalization strategies: Introduction of VLMs like CLIP and DINOv2 offers a semantic-level generalization pressure mechanism. More effective leveraging of these multimodal pretrained knowledge for UAD may reduce dependence on labeled spoof data.

  3. Task protocol and standard reconstruction: Existing protocols, though representative, struggle to cover mixed attacks, multimodal scenarios, and mobile deployment. Developing higher-level task definitions and layered evaluation mechanisms (e.g., distinguishing high-risk from tolerable errors) is necessary to enhance real-world reliability.

Currently, we still walk within the mist of recognizing spoofs.

Different teams have illuminated various algorithmic paths—some piercing anomalies via attention, some building robustness through masking, some disentangling spatial truth via 3D reconstruction.

But the ultimate question remains unchanged:

Do we really understand how a fake face is composed?

Perhaps the answer is yet to form, but after this open technical contest, we have stepped one step closer.