acceptodds
Under review as a conference paper at ICLR 2027

EviTrace: Structured Evidence-Guided Progressive Reasoning for Interpretable Face Anti-Spoofing

Abstract

Existing face anti-spoofing methods typically apply a uniform one-pass pipeline to all samples, despite the substantial variation in the saliency and diagnostic difficulty of spoofing cues. While some attacks can be identified from obvious artifacts with minimal reasoning, others exhibit only subtle and hard-to-perceive clues that require more careful evidence aggregation. To address this limitation, we propose EviTrace, a structured Evidence-guided progressive reasoning framework for interpretable face anti-spoofing. Specifically, we first introduce Structured Evidence Slots (SES), which leverage structured priors during training to learn explicit responses to diverse spoof-related cues beyond sample-level binary live/spoof supervision. Building on these structured evidence responses, we further develop Difficulty-Aware Progressive Reasoning (DPR), which pools high-confidence slot evidence for coarse screening and selectively routes low-confidence samples for refinement by revisiting the most ambiguous slots and their most relevant local patch regions. Rather than replacing the coarse prediction, DPR progressively corrects it through residual evidence aggregation. Based on this evidence-to-decision process, EviTrace naturally yields an Interpretable Reasoning Trace (IRT) that explicitly reveals evidence activation, uncertainty estimation, selective refinement, and final correction as a transparent decision chain. Experiments under a generalization setting show that EviTrace achieves competitive performance while providing a clearer and more transparent reasoning process. Code is available at https://anonymous.4open.science/r/EviTrace

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Face Anti-Spoofing, Evidence-Guided Reasoning, Progressive Reasoning, Difficulty-Aware Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_X1GMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EviTrace uses spoof-type-supervised evidence slots and confidence-guided residual refinement for face anti-spoofing. A coarse prediction routes uncertain samples to selected slots and patches, exposing an explicit intermediate trace. Training uses CelebA-Spoof, and evaluation covers eleven target datasets with component ablations, slot suppression, and measured inference costs.

Critical strengths
  1. The intermediate variables are connected to the actual prediction path. Slot suppression and routing/patch ablations provide more meaningful evidence than visualizations alone, while the residual formulation preserves a concrete coarse-to-fine decision process.
  2. The evaluation covers diverse unseen domains and separates discriminative from generative comparators. The appendix also reports actual mixed-batch latency, making the limits of selective-computation savings explicit.
Critical weaknesses
  1. The trace exposes auditable computation, but does not yet provide a validated explanation of spoof cues. Slots are supervised by attack-type labels, and suppressing a predictive slot establishes functional relevance rather than semantic correctness or localization faithfulness. The paper acknowledges this distinction, which still limits its central interpretability claim.
  2. The richer-supervision control uses a plain ViT-B, whereas EviTrace uses a CLIP-pretrained backbone. It therefore does not fully isolate the contribution of structured slots from a conventional CLIP-based classifier with the same ten-way supervision and training setup.
Questions for authors
  1. How does EviTrace compare with a CLIP-based ten-way auxiliary classifier trained with the same labels, sampler, and optimization budget?
  2. Can sample-level cue annotations or targeted region interventions validate slot semantics and the relevance of DPR-selected patches?
3 comments

[Q1] Interpretability of the reasoning trace

Thank you for raising this concern. We agree that IRT provides an auditable inference trace, but attack-type supervision alone does not establish the semantic correctness of individual spoof cues. To further assess its interpretability, we conducted matched slot interventions on SiW to evaluate whether the learned evidence slots exhibit attack-family-level semantic selectivity.

Attack-family selectivity of the slots. We grouped the annotated attacks into Print and Replay. We suppressed either the three print slots (Photo, Poster, and A4) or the three display slots (PC, Pad, and Phone), using the same suppression procedure, checkpoint, test videos, and decision threshold. We report APCER, the fraction of attack videos incorrectly accepted as bona fide, separately for each attack family.

Ground-truth attack familyOriginal APCER (%) ↓Suppress print slots (%) ↓Suppress display slots (%) ↓
Print1.403.261.63
Replay1.591.594.16

For Print attacks, suppressing the print slots increases APCER from 1.40% to 3.26%, compared with 1.63% when suppressing display slots. For Replay attacks, suppressing the display slots increases APCER from 1.59% to 4.16%, compared with 1.59% when suppressing print slots. The reversal across the two attack families supports semantic selectivity at the attack-family level. Each slot group has a larger effect on the detection of its corresponding attack family. The functional relevance of DPR-selected slots and patches is further evaluated through matched removal experiments, as detailed in our response to Q3.

[Q2] Matched supervision for structured evidence slots

Thank you for raising this point. The original plain ViT-B control was designed to assess the benefit of adding ten-way attack-type supervision on a fixed backbone. To isolate the contribution of the slot design under matched CLIP pretraining, we added a conventional auxiliary head that predicts the ten attack types from the global CLS feature alongside the binary live/spoof task.

The CLS auxiliary-head and slot-supervised variants use the same CLIP ViT-B/16 initialization, ten attack-type labels, sampling strategy, and optimization settings. Both use a type-loss weight of 0.15, with slot–text alignment and DPR disabled.

CLIP ViT-B/16 variantAvg. HTER (%) ↓Avg. AUC (%) ↑
Binary classification only21.0185.68
Binary classification + evidence slot tokens21.3185.43
Binary classification + CLS auxiliary type head22.6183.96
Evidence slot tokens + slot-wise type supervision15.1389.43

Across the 11 target datasets, slot-wise supervision reduces average HTER from 22.61% to 15.13% and increases average AUC from 83.96% to 89.43% compared with the CLS auxiliary head. Adding slot tokens alone yields performance close to that of the binary baseline. The gains emerge when the attack-type labels supervise the slots.

One possible explanation is that the auxiliary head places two demands on the same global CLS feature. The type classifier uses it to distinguish among attacks, while the binary classifier uses it to group those attacks into a single spoof class. Dedicated slots give the model a place to learn distinct attack cues and combine their pooled representations with the global CLS feature for the binary decision. The matched results support the benefit of organizing and using the same attack-type supervision through dedicated evidence slots.

We will add the matched CLS auxiliary-head comparison to Table 2 and clarify the purpose of the original plain ViT-B control in the revised manuscript.

[Q3] Validation of slot semantics and DPR-selected patches

Thank you for suggesting these validation tests. As shown in our response to Q1, the matched Print/Replay interventions support attack-family-level semantic specificity. Here, we further evaluate the functional relevance of the slots and patches selected by DPR through equal-count removal experiments.

Functional relevance of DPR-selected patches. For the patch intervention, we fixed the slots used to retrieve patches and removed either the 49 patches originally selected by DPR or 49 randomly sampled patches from the same candidate pool. DPR then retrieved 49 patches from the remaining candidates using the original affinity rule. We also ran an analogous comparison for the four slots selected by DPR. All conditions used the same checkpoint, coarse predictions, routed samples, and decision threshold. HTER was computed over the full SiW test set.

Inference conditionNumber removedHTER (%) ↓Absolute HTER increase (%)
Original EviTrace01.690.00
Remove random slots43.03+1.34
Remove DPR-selected slots46.36+4.67
Remove random patches492.05+0.36
Remove DPR-selected patches495.09+3.40

Removing DPR-selected patches increased HTER by 3.40% in absolute terms, compared with 0.36% for equal-count random removal. Because DPR used the same number of patches in both conditions, this difference supports the contribution of the particular regions selected by DPR to the final decision. The slot intervention supports attack-family semantics, and the patch intervention supports the functional relevance of the selected regions. Sample-level annotations of specific cues and their locations would enable a finer-grained test of whether individual slots capture those cues and whether DPR localizes them.

Reviewer_wRDhOpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EviTrace, an interpretable FAS method guided by structured evidence. EviTrace consists of two main components, SES and DPR, which learn explicit responses to different spoofing cues and adaptively process samples of varying difficulty. The method is evaluated under the widely adopted CelebA-Spoof 1-to-11 cross-domain evaluation protocol and achieves competitive results. The authors also conduct ablation studies to validate the effectiveness of the proposed components.

Critical strengths
  1. The proposed method leverages different spoofing cues as additional supervision signals to enhance the model's discriminative capability.

  2. The DPR module is carefully designed, incorporating both confidence-based routing and correction strength control to adaptively combine predictions from the two reasoning stages.

  3. Despite relying only on a CLIP-based backbone, the proposed method outperforms the VLM-based I-FAS method and achieves competitive performance. The authors also provide some discussion of the method's limitations.

Critical weaknesses
  1. In the case studies presented in Figure 4, the authors should provide a more detailed analysis of the selected patches for different spoofing samples. It is difficult to intuitively identify the corresponding spoofing types or understand why these particular patches are considered critical for the final predictions.

  2. The third paragraph of the Introduction suffers from a confusing narrative structure and lacks a clear logical flow. The relationships between the presented ideas should be better organized and clarified.

  3. The evidence slots in SES are directly defined based on the 10 spoofing types annotated in CelebA-Spoof. Such a predefined, source-domain-specific slot design may inherently limit the model's ability to detect novel spoofing types that are not covered by these categories.

Questions for authors
  1. Address the first two weaknesses.

  2. The proposed method achieves competitive results with CLIP. Why not extend it to MLLMs to explore whether it can achieve greater improvements?

6 comments

[Q1] Case analysis of Figures 4 and 6

Thank you for asking us to explain why DPR selects these patches and how they affect the final decision. When the coarse prediction has low confidence, DPR uses the least-confident evidence slots to retrieve patches with high slot–patch affinity. It then aggregates the retrieved features to compute a residual update to the coarse score. The table summarizes the selected regions and score changes in both figures.

CaseExamples of selected regionsSpoof probability before → after DPRPrediction change
Fig. 4, top leftForehead, eye area, nose bridge, and cheeks0.542 → 0.426Spoof → Live
Fig. 4, top middleEyes, nose, and adjacent cheeks0.606 → 0.484Spoof → Live
Fig. 4, top rightCheeks, nose, and lower face0.574 → 0.400Spoof → Live
Fig. 4, bottom leftEye area, cheeks, and jawline0.470 → 0.512Live → Spoof
Fig. 4, bottom middleRegions around the glasses, cheeks, and lower face0.477 → 0.562Live → Spoof
Fig. 4, bottom rightNose, cheeks, and mouth area0.445 → 0.547Live → Spoof
Fig. 6, leftImage borders and the screen area behind the subject0.478 → 0.581Live → Spoof
Fig. 6, middleCap, background, and clothing0.577 → 0.342Spoof → Live
Fig. 6, rightWall, image margins, and clothing0.473 → 0.577Live → Spoof

Successful refinement in Figure 4. The top-middle bona fide example illustrates how neighboring patches around the eyes, nose, and cheeks can provide complementary information about local texture and transitions between facial structures. Stage I summarizes evidence across the face crop, while DPR gives these retrieved features a direct role in the score update. In this example, the update lowers the spoof probability enough to correct the decision. The bottom-right attack example shows why spatial location alone is insufficient to interpret the update. DPR retrieves patches from overlapping facial areas, but their contribution raises the spoof score. The opposite changes illustrate that the score update depends on the features in the retrieved regions and their interaction with the selected slots. Our matched removal experiment on SiW supports the contribution of these selections at the dataset level. Removing DPR-selected patches raises HTER to 5.09%, compared with 2.05% after removing the same number of random patches.

Failed refinement in Figure 6. In the rightmost bona fide example, many selected patches fall on the wall or near the image margins, leaving relatively little coverage of the transitions between the nose and cheeks. Strong background edges can receive high affinity scores and occupy part of the fixed patch budget. With fewer complementary facial features available, the residual update may give greater weight to misleading context and raise the spoof probability enough to reverse the correct coarse decision. In the middle attack example, patches on the cap, background, and clothing occupy part of the selected set, leaving limited coverage of the central face. One plausible explanation is that the retrieved features provide weak support for the original attack decision, and the residual update lowers the spoof probability. These cases illustrate how the retrieved evidence can move a low-confidence prediction in the wrong direction.

We will expand Section 3.5 to connect the selected regions, the score changes before and after refinement, and the final decisions in both figures. We will also add ground-truth attack subtypes where the dataset annotations provide them.

[Q2] Logical flow of the third paragraph of the Introduction

Thank you for pointing out the unclear flow in the third paragraph. We reorganized it to connect the supervision problem to the design of our two-stage detector. The revised paragraph first explains why binary labels leave individual spoofing cues implicit. It then explains how explicit evidence slots support the initial decision and help identify ambiguous evidence. This leads to the two decisions made during inference. Coarse prediction confidence determines whether to refine, and slot confidence guides which local regions to revisit.

Revised paragraph

Binary supervision provides a live/spoof label for each sample, leaving the individual cues behind that label implicit. Representing these cues explicitly allows the model to aggregate them for an initial decision and identify which pieces of evidence remain ambiguous. This motivates a two-stage detection strategy. The model accepts a sufficiently confident coarse prediction directly. When the coarse prediction is uncertain, the model uses the least-confident evidence slots to retrieve relevant local image regions and refine the decision.

[Q3] Predefined evidence slots and unseen attack types

Thank you for raising this concern. We agree that defining the slots from source-domain attack categories can limit generalization to genuinely novel attacks, especially when their visual cues are poorly represented in the source data. The ten attack types annotated in CelebA-Spoof cover a broad range of common physical presentation attacks, including print, replay, and mask attacks. This makes CelebA-Spoof a practical starting point for learning structured evidence. EviTrace reuses the dataset’s existing attack-type labels to learn distinct slot responses and guide local refinement, so the same coarse labels support both evidence learning and evidence-guided detection.

An unseen presentation medium can still share visual cues with known attacks. For example, a novel print material may exhibit a planar facial appearance or texture distortions that the model has learned from other print attacks. EviTrace combines evidence from multiple slots with global image features to make its binary live/spoof decision. This aggregation offers a potential benefit for unseen attacks. When a new attack shares cues with several source categories, multiple slots can contribute complementary evidence to the same decision. How well this transfers depends on the visual cues present in the new attack and how well the source data cover them.

Attack-type labels are used only during training, and inference requires only the input image. As more comprehensive training datasets become available, their coarse attack-type labels can be used to extend the slot definitions and retrain EviTrace within the same framework. Text alignment uses fixed, category-specific templates based on these labels, keeping annotation at the attack-type level. We will clarify this potential benefit of combining slot evidence, along with its dependence on source-domain coverage, in the revised manuscript.

Reviewer_Yfa4OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses two problems in face anti-spoofing (FAS): cross-domain generalization and the explainability of detection decisions. The basic idea is Structured Evidence-Guided Progressive Reasoning. The proposed method, EviTrace, introduces Structured Evidence Slots (SES) supervised with 10 attack types on top of CLIP ViT-B/16, and combines them with Difficulty-Aware Progressive Reasoning (DPR) to refine predictions. With this, the method provides an Interpretable Reasoning Trace (IRT) that tracks the slots and regions used for the decision, along with the scores before and after refinement. CelebA-Spoof is used for training, and evaluation is done on 11 target datasets. Across the 11 targets, EviTrace has the best reported average performance among discriminative methods. Among MLLM-based methods, it does better than I-FAS but is generally below FaceCoT and TAR-FAS. In the evaluation, the authors also present component ablations, hyperparameter sensitivity, slot response analysis, counterfactual slot suppression experiments, and inference efficiency measurements.

Critical strengths
  • The paper addresses an important problem. Cross-domain generalization and the explainability of detection decisions in face anti-spoofing (FAS) are both areas that need further research.
  • The authors put effort into a broad cross-domain evaluation. It’s good to see an average HTER of 9.98% on 11 unseen datasets, along with results from repeated runs.
  • The paper reports the reasoning process and its measured cost. In particular, it gives intermediate decision records together with latency. It also reports about 8.7 ms of model computation at batch size 1 on an RTX 4090, so practicality can be checked as well.
  • The code is also made available.
Critical weaknesses

The motivation to improve generalization and explainability together is reasonable. However, the Related Work section is placed entirely in the appendix, so it’s hard to tell from the main text how the method differs from key prior work. The bigger problem is that closely related prior work is missing. In the main text, the discriminative ViT and vision-language baselines in Table 1 are limited to ViTAF, ViT, and FLIP. At a minimum, comparisons with directly related methods such as MVP-FAS (ICCV’25), CFPL-FAS (CVPR’24), and BUDoPT (ECCV’24) are required. On the technical side, the paper should also explain more clearly how the motivation of learning physical spoofing cues relates to the implementation, which in practice maps attack-type labels to slots. More details follow.

  • Closely related prior work and recent baselines are lacking. In particular, comparisons with directly related methods such as MVP-FAS (ICCV’25), CFPL-FAS (CVPR’24), and BUDoPT (ECCV’24) are missing, so I can’t judge the novelty or the performance advantage over discriminative models. Because the comparisons in the main-text tables were limited, I looked for the Related Work section and found it in the Appendix. I understand the choice to put it in the Appendix due to space limits, but even there the range of papers covered is limited. Leaving out methods that must be compared against is a serious oversight.
  • On the technical side, the core contributions aren’t validated in isolation. The paper reports an overall performance gain but doesn’t sufficiently isolate its cause. Table 2 adds DPR only after SES is in place, so there’s no comparison with DPR applied alone without SES, and there’s no analysis of whether the routed samples are in fact hard samples. As a result, the current experiments don’t support the claim that the gain comes from evidence-guided, difficulty-aware reasoning, which is the paper’s core contribution. To check clearly whether the effect comes from SES or from the extra supervision, the paper must include a control baseline that adds only auxiliary attack-type classification to the same CLIP model and uses matched sampling.
  • Text alignment improves HTER by 1.57 percentage points (pp) in the SES configuration of Table 2. However, in the full model in Table 8, removing it changes the result by 0.06 pp, a negligible amount. Of course, these are different settings, so the two results don’t contradict each other. Still, the independent contribution of semantic alignment in the full model is unclear. The paper also needs to explain the results for each loss weight and report repeated runs.
  • The comparisons in Tables 3 and 4 should also be made under matched pre-refinement performance, trained model, and threshold settings. Otherwise the effect of each stage can’t be judged clearly. It’s also unclear how the development split is built for each target dataset and what data were used to select the model and hyperparameters.
  • The computational efficiency claim should be read with caution. The backbone accounts for about 99.8% of the total computation, and at large batch sizes the latency benefit of selective routing isn’t clear either.

[1] MVP-FAS Multi-View Slot Attention Using Paraphrased Texts for Face Anti-Spoofing ICCV 2025.

[2] CFPL-FAS CFPL-FAS: Class Free Prompt Learning for Generalizable Face Anti-spoofing CVPR 2024.

[3] BUDoPT Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing ECCV 2024

Questions for authors
  • Are any of the points I listed under Weaknesses incorrect?
  • Under the same evaluation and supervision conditions, does the improvement still hold against representative recent methods such as MVP-FAS (ICCV’25), CFPL-FAS (CVPR’24), and BUDoPT (ECCV’24)?
  • Does SES outperform a control baseline that adds auxiliary attack-type classification to the same CLIP model with matched sampling? If the checkpoint and the decision threshold are selected using only validation data from the source domain (i.e., CelebA-Spoof), what are EviTrace’s HTER and AUC on the 11 targets? Under this setting, can you also show whether EviTrace still improves over SES-only and the EviTrace binary baseline, and how the routing ratio varies across targets?
  • In the full model, removing text alignment changes the result by 0.06 pp. Do you consider this a meaningful difference or a negligible one? What is the basis for that judgment?
  • Can you provide quantitative evidence that the selected slots and regions reflect real attack evidence? I’d like to know whether removing them affects the prediction more than removing the same number of random slots or regions, and if so, by how much.
13 comments

[Q1] Closely related work and recent baselines

Thank you for pointing out the missing comparisons. Since these methods did not report results under our One-to-Eleven evaluation protocol, they were not included in the original submission. We have now reproduced their results under the same benchmark and expanded the Introduction, Related Work, experimental comparisons, and results analysis to incorporate these representative methods.

We added MVP-FAS (ICCV 2025), CFPL-FAS (CVPR 2024), BUDoPT (ECCV 2024), and ICPE-FAS (IJCV 2026), the journal extension of CFPL-FAS, to Table 1. The revised table reports results on all 11 target datasets. We used the official implementations of MVP-FAS, CFPL-FAS, and ICPE-FAS and reimplemented BUDoPT from the published paper.

MethodAvg. HTER (%) ↓Avg. AUC (%) ↑
BUDoPT (ECCV 2024)20.3486.86
MVP-FAS (ICCV 2025)17.8289.00
CFPL-FAS (CVPR 2024)14.3291.66
ICPE-FAS (IJCV 2026)12.5492.39
EviTrace9.9894.58

EviTrace achieves the lowest average HTER and highest average AUC among the compared discriminative methods. Relative to CFPL-FAS, EviTrace reduces average HTER by 4.34% and increases average AUC by 2.92% in absolute terms. Relative to ICPE-FAS, the strongest added baseline, the corresponding gains are 2.56% and 2.19%.

We expanded the Introduction and Related Work to clarify how EviTrace differs from each baseline. BUDoPT uses domain prompts to account for variations in recording conditions and attack types. EviTrace addresses the uncertainty that remains on individual samples by using coarse confidence to decide when further analysis is needed. CFPL-FAS uses instance-conditioned prompts to reweight visual features. EviTrace learns dedicated attack-type slots so that specific evidence can be revisited when it is ambiguous. ICPE-FAS also uses category-level prompts to construct classifier weights and suppress domain-related signals. EviTrace can update an uncertain prediction directly through a residual correction informed by local evidence. MVP-FAS employs paraphrased live/spoof descriptions to guide multi-view slot attention and multi-text patch alignment, primarily enhancing representation learning. In contrast, EviTrace explicitly supervises attack-type-specific slots and integrates their evidence into the initial prediction. For uncertain samples, it further selects the least-confident slots, retrieves their most relevant local patches, and performs residual score correction. Thus, EviTrace extends structured evidence learning to difficulty-aware, selective refinement at inference time.

EviTrace uses supervised attack-type slots to inform the initial prediction and guide a second pass when that prediction is uncertain. For each routed sample, DPR revisits the least-confident slots and uses their affinity to local patches to retrieve evidence for a residual score update. The main-text ablations show how these steps contribute to the average performance. In Table 2, adding slot supervision reduces HTER from 21.31% to 15.13%, and adding DPR to the full SES model further reduces HTER from 13.56% to 9.98%. In Table 3, replacing the evidence-guided patches with random patches raises HTER to 10.85%. These results support the contribution of both learning attack-specific evidence and using it to refine uncertain predictions.

** [Q2-1] Component ablations and source-only generalization**

(1) Evidence slots versus a conventional auxiliary head. We added a ten-class attack-type head on the global CLS feature alongside the binary classifier. Both variants use identical CLIP ViT-B/16 initialization, attack-type labels, sampling, optimization, and type-loss weight (0.15), with slot–text alignment and DPR disabled.

CLIP ViT-B/16 variantAvg. HTER (%) ↓Avg. AUC (%) ↑
Binary classification only21.0185.68
Binary classification + CLS auxiliary type head22.6183.96
Evidence slot tokens + slot supervision15.1389.43

The slot-supervised variant reduces HTER by 7.48% and improves AUC by 5.47% over the matched auxiliary-head baseline. Unlike the auxiliary head, which supervises the shared CLS representation, EviTrace assigns type supervision to dedicated slots and aggregates their evidence for binary detection. This supports the benefit of structured evidence organization beyond additional supervision alone.

(2) DPR without explicit slot semantics. Since DPR requires slot features for local patch retrieval, we retained learnable slot tokens but disabled attack-type supervision and slot–text alignment. These slots learn through binary supervision alone. We then evaluated DPR with and without full SES.

VariantExplicit slot supervision and text alignmentDPRAvg. HTER (%) ↓Avg. AUC (%) ↑
Slot tokens only——21.3185.43
Slot tokens + DPR—✓18.8887.71
SES-only✓—13.5692.15
Full EviTrace✓✓9.9894.64

Without explicit slot semantics, DPR reduces HTER by 2.43% and improves AUC by 2.28%. With full SES, these gains increase to 3.58% and 2.49%, respectively. Thus, DPR improves both configurations, with greater gains from semantically supervised slots. Table 3 further supports evidence-guided selection: replacing selected patches with random patches increases HTER from 9.98% to 10.85%.

(3) Difficulty of routed samples before refinement. Using Stage I predictions from the same checkpoint, we applied identical live/spoof thresholds to routed and unrouted samples within each target. Coarse error rate measures the proportion misclassified by Stage I.

GroupCoarse error rate (%)
Routed, q<0.45q<0.457.88
Unrouted, q≥0.45q\geq0.455.50

Routed samples have a 2.38% higher coarse error rate, confirming that routing prioritizes samples more likely to be misclassified. DPR also achieves positive net correction in both routed confidence ranges: +5.75% for q<0.225q<0.225 and +4.34% for 0.225≤q<0.450.225\leq q<0.45.

[Q2-2] Comparable target-domain data access

Thank you for examining the evaluation protocol closely and highlighting the role of target-domain labels in both checkpoint selection and threshold calibration. In face anti-spoofing, cues from physical presentation attacks can be subtle, while differences in sensors, illumination, compression, and attack media can substantially shift the score distributions of bona fide and spoof samples. A detector can separate the two classes on a new target domain even when a threshold calibrated on the source domain yields a poor operating point. Source-domain checkpoint selection can also favor features tied to the source capture conditions or attack media. Our protocol therefore evaluates cross-domain transfer in a setting where labeled target development data are available to select a checkpoint and calibrate the operating point.

Unified domain generalization setting. We update all model weights using only the official CelebA-Spoof training split. We select one shared checkpoint by minimizing macro-average HTER across the 11 target development sets. We then calibrate a separate live/spoof threshold on each target development set and apply it to the corresponding held-out test set. Target development labels thus inform both checkpoint selection and threshold calibration. We confirmed that the principal baselines in Table 1, including FLIP (ICCV 2023), BUDoPT (ECCV 2024), CFPL-FAS (CVPR 2024), I-FAS (AAAI 2025), MVP-FAS (ICCV 2025), FaceCoT (CVPR 2026), TAR-FAS (CVPR 2026), and ICPE-FAS (IJCV 2026), follow the same source-training and target-domain calibration strategy as EviTrace. These methods train on labeled source-domain data and use labeled target-domain data to calibrate the live/spoof decision threshold for target-domain evaluation. The released implementations of earlier methods, including SSDG (CVPR 2020), SSAN (CVPR 2022), and SA-FAS (CVPR 2023), also use labeled target-domain data to determine the operating point.

Under the common protocol, EviTrace’s lower HTER and higher AUC relative to the reproduced baselines support better cross-domain detection and score separation when labeled target development data are available. We agree that deployment without target labels is a stricter setting, in which both the checkpoint and a fixed decision threshold must be selected using source-domain data alone.

Source-only domain generalization. Following the reviewer’s suggestion, we also trained on the official CelebA-Spoof training split and used only the CelebA-Spoof validation split for checkpoint selection and live/spoof threshold calibration. For each method, we fixed the selected checkpoint and its calibrated threshold across all 11 target datasets. We used target labels only to compute the test metrics. The table reports macro-averaged results across the 11 targets.

MethodAvg. HTER (%) ↓Avg. AUC (%) ↑
EviTrace binary baseline46.2276.02
EviTrace with SES only28.2483.81
Full EviTrace (SES + DPR)23.9587.75
MVP-FAS (ICCV 2025)32.4287.58
BUDoPT (ECCV 2024)38.9076.82

Under source-only checkpoint selection and threshold calibration, adding SES to the binary baseline reduces average HTER from 46.22% to 28.24% and raises average AUC from 76.02% to 83.81%. Adding DPR further improves both metrics, bringing HTER to 23.95% and AUC to 87.75%. Full EviTrace also achieves lower HTER than MVP-FAS (32.42%) and BUDoPT (38.90%). Its AUC is close to that of MVP-FAS (87.58%) and higher than that of BUDoPT (76.82%). These results show that both SES and DPR improve cross-domain detection when model selection and threshold calibration rely solely on source-domain validation data.

Comments on the paper

9 comments

[Q1] Matched supervision

Thank you for raising this important question. The original plain ViT-B control examined the benefit of additional attack-type supervision without CLIP pretraining, but did not isolate the contribution of SES under matched conditions. Following the reviewer's suggestion, we added a conventional auxiliary type classifier to evaluate whether the structured organization of attack-type annotations provides benefits beyond richer supervision alone.

Specifically, the conventional baseline predicts ten attack types from the global CLS feature alongside the binary live/spoof classifier. Both the CLS auxiliary-head and slot-supervised variants use identical CLIP ViT-B/16 initialization, attack-type labels, sampling strategy, and optimization settings, with the type-supervision weight fixed at 0.15. Slot-text alignment and DPR are disabled to isolate the effect of structured slot supervision.

CLIP ViT-B/16 variantAvg. HTER (%) ↓Avg. AUC (%) ↑
Binary classification only21.0185.68
Binary classification + CLS auxiliary type head22.6183.96
Evidence slot tokens only21.3185.43
Evidence slot tokens + slot supervision15.1389.43

Experimental results. Adding a conventional CLS auxiliary type head degrades HTER/AUC from 21.01/85.68 to 22.61/83.96, indicating that additional attack-type supervision does not necessarily improve cross-domain detection. Similarly, introducing evidence slot tokens without supervision provides no benefit (21.31/85.43). In contrast, applying structured supervision to these slots substantially improves performance to 15.13/89.43, reducing HTER by 7.48% and increasing AUC by 5.47% over the matched CLS auxiliary-head baseline.

Mechanism analysis. The explanation is that conventional auxiliary classification imposes both fine-grained attack-type discrimination and binary detection on the same CLS representation, potentially introducing conflicting objectives under domain shift. In contrast, SES assigns type-specific supervision to dedicated evidence slots, which interact with the CLS token and image patches through Transformer self-attention, allowing attack-specific cues to inform the global representation. The learned slot evidence is further aggregated and combined with the CLS feature for binary prediction. This enables structured evidence learning while facilitating information exchange between type-specific and global representations.

These matched comparisons suggest that the improvement arises not simply from richer supervision or additional tokens, but from how SES organizes, exchanges, and integrates type-specific evidence for binary detection. We will incorporate this control into Table 2 and clarify its distinction from the original plain ViT-B experiment in the revised manuscript.

[Q2] Semantic specificity

Thank you for suggesting this control. We agree that IRT exposes the computational provenance of the explicit evidence-selection and refinement process, rather than providing a complete semantic decomposition of all information used by the classifier. The CLS feature and high-dimensional slot representations also contribute to Eqs. (7) and (12), while the displayed slot responses and selected regions expose only the corresponding interpretable variables. Table 5 establishes the functional relevance of these variables, but does not by itself show whether a named slot is selectively associated with its stated attack semantics.

To evaluate this semantic specificity, we conducted a matched intervention on two attack families in SiW. Following its original attack categories, we grouped the attacks into Print and Replay. We compared two semantically defined slot groups:

  • Print slots: Photo, Poster, and A4.
  • Display slots: PC, Pad, and Phone.

Each test video was evaluated under three conditions: original inference, suppression of the three print slots, and suppression of the three display slots. All conditions used the same checkpoint, test videos, and decision threshold. Both interventions used the same slot-suppression operation as Table 5 and removed the same number of slots, ensuring that their effects differed only in semantic correspondence.

We report APCER, which measures the percentage of attack videos incorrectly accepted as bona fide. While the HTER reported in Table 5 summarizes errors over both attack and bona fide samples, APCER directly measures the effect of each intervention on the corresponding attack family.

Ground-truth attack familyOriginal APCER (%)Suppress print slots (%)Suppress display slots (%)
Print1.403.261.63
Replay1.591.594.16

For Print attacks, suppressing the semantically corresponding print slots increases APCER from 1.40% to 3.26%, an absolute increase of 1.86 percentage points. Suppressing the non-corresponding display slots increases APCER by only 0.23 percentage points. The corresponding intervention therefore causes 1.63 percentage points more degradation than the non-corresponding intervention.

For Replay attacks, suppressing the corresponding display slots increases APCER from 1.59% to 4.16%, an absolute increase of 2.57 percentage points. Suppressing the non-corresponding print slots leaves APCER unchanged at 1.59%. The corresponding intervention therefore causes 2.57 percentage points more degradation.

This crossover provides evidence for semantic specificity at the attack-family level. The same slot group has a substantially larger effect when its corresponding attack family is present, whereas suppressing a non-corresponding group has little effect. Together with Table 5, these results show that the named slots are not merely useful intermediate features. Their influence on the decision is selectively associated with the corresponding attack families.

[Q3] Refinement utility

Thank you for asking which frames benefit from DPR and whether coarse confidence helps allocate a fixed refinement budget. We use coarse and refined predictions from the same EviTrace checkpoint. For each dataset, we apply the same live/spoof decision threshold before and after refinement and across routing policies. W→C and C→W denote the percentages of refined frames in each row whose predictions change from wrong to correct and from correct to wrong, respectively. Net correction is W→C minus C→W.

Refinement across confidence ranges. We split the frames routed to DPR at q=0.225q=0.225, the midpoint of the routing interval 0≤q<0.450\le q<0.45.

Coarse-confidence rangeW→C (%) ↑C→W (%) ↓Net correction (%) ↑
Lower confidence, 0≤q<0.2250\le q<0.2255.920.17+5.75
Closer to the routing threshold, 0.225≤q<0.450.225\le q<0.454.420.08+4.34

DPR corrects more errors than it introduces in both ranges. The lower-confidence group achieves an absolute net correction gain of 1.40% over the group closer to the routing threshold. This difference comes mainly from a higher rate of corrected coarse errors (5.92% vs. 4.42%), with similar rates of harmful changes (0.17% vs. 0.08%).

Equal-budget routing comparison. We retain EviTrace’s original routing rule (q<0.45q<0.45). On each dataset, the random policy draws exactly as many test frames as EviTrace routes, uniformly without replacement. Both policies apply the same DPR refinement branch to the selected frames and retain the coarse predictions for the remaining frames.

Routing policyRefined frames (%)W→C (%) ↑C→W (%) ↓Net correction (%) ↑
Uniform random routing≈16.751.670.24+1.43
EviTrace confidence-based routing (Table 3)≈16.755.140.12+5.02

At the same refinement budget, confidence-based routing achieves 5.02% net correction, compared with 1.43% for uniform random routing, an absolute gain of 3.59%. The gain comes primarily from correcting more coarse errors (5.14% vs. 1.67%), with fewer correct decisions changed to incorrect ones (0.12% vs. 0.24%). With the refinement branch and decision threshold held fixed, this result supports using coarse confidence to allocate the refinement budget.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.