acceptodds
Withdrawn from ICLR 2027

Calibrate Elsewhere, Verify Here:Disjoint In-Scene Calibration for Multimodal Face Anti-Spoofing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal face presentation attack detection must distinguish presentation effects from ordinary changes in illumination and sensing. Target-only inference can marginalize an unknown acquisition state or estimate it from the face; the former adds no sample-specific information about the current capture, while the latter can suppress nuisance-aligned attack evidence. We introduce SceneCal, which learns type-conditioned responses of heterogeneous non-target regions to a shared acquisition state. Controls determine a posterior that is fixed before the face is scored, with uncertainty propagated to partially observed target modalities. We analyze plug-in attenuation, target-only marginalization, and disjoint control conditioning in one Gaussian model. On public WMCA RGB–Depth leave-one-attack-out, SceneCal reaches 2.99% ACER, versus 3.37% for CTNet and 3.54% for MMDG++; on HQ-WMCA it reaches 9.21%, versus 9.88% and 10.23%. On HQ-MMFAS, CTNet is stronger on known instruments (2.06% vs. 3.08%), while SceneCal reaches 4.72% on held-out instruments, compared with 5.08% for CTNet and 5.45% for DADM.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal representation learning, cross-modal learning, intra-sample generalization, authenticity verification, face anti-spoofing, anomaly detection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uGbsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SceneCal, a multimodal face presentation attack detection framework that separates acquisition-state calibration from face verification. Instead of estimating acquisition conditions directly from facial observations, SceneCal leverages heterogeneous non-target regions to infer a shared latent acquisition state and its posterior uncertainty. The inferred state is then used to construct a conditional predictive distribution of bona-fide facial responses, enabling attack detection through uncertainty-aware residual scoring. The framework incorporates type-conditioned cross-modal modeling, robust control inference, and bona-fide-only training. The paper also provides a theoretical analysis comparing target-only plug-in estimation, marginalization, and disjoint control conditioning under a linear-Gaussian model. Experiments on WMCA, HQ-WMCA, and HQ-MMFAS evaluate the method under known and unseen attack settings, supported by ablation studies, cross-dataset transfer, and robustness analyses. SceneCal demonstrates improved performance in several unseen-attack scenarios, while existing supervised approaches remain competitive on known attacks.

Critical strengths

<1>The paper addresses an important challenge in multimodal face anti-spoofing: distinguishing acquisition-related variations from presentation attack evidence. Instead of directly adapting to facial observations, SceneCal estimates the shared acquisition state from disjoint non-target regions and incorporates posterior uncertainty into face verification. This formulation provides a principled way to mitigate the ambiguity between acquisition variations and attack-induced deviations, making the overall methodological framework meaningful and well-motivated.

<2>The paper provides a coherent theoretical analysis of target-only adaptation, marginalization, and disjoint control conditioning under a unified probabilistic framework. The experimental evaluation is relatively comprehensive, including comparisons with recent methods, mechanism analyses, ablation studies, cross-dataset transfer, and robustness evaluations. In particular, the results on unseen attacks and the control-region experiments provide useful evidence supporting the proposed calibration mechanism. These analyses strengthen the credibility of the reported improvements and help explain when and why the method is effective.

Critical weaknesses

<1>Although the proposed disjoint calibration framework is well-motivated, its main technical components, including latent Gaussian modeling, posterior inference, negative-control-based nuisance estimation, and robust reweighting, are largely based on established statistical techniques. The paper does not sufficiently clarify the methodological novelty of combining these components beyond their application to multimodal face anti-spoofing. Furthermore, the improvements over strong recent baselines are relatively modest, particularly in unseen-attack settings. This raises concerns about whether the proposed framework provides a sufficiently substantial methodological contribution relative to its additional modeling complexity.

<2>The proposed framework assumes that non-target control regions share acquisition factors with the face while remaining unaffected by the presentation attack mechanism. Although this assumption may be reasonable for localized attacks, such as facial masks, it may not hold for large-screen replay attacks or scenarios where presentation devices introduce spatially correlated illumination or reflection changes across the scene. In such cases, multiple control regions may be systematically affected by the attack, potentially biasing acquisition-state estimation and suppressing attack-related evidence. While the authors investigate control contamination using robust IRLS, the current experiments do not sufficiently establish whether the method remains reliable under such structured, attack-induced contamination. This limitation raises concerns about the practical applicability and robustness of the proposed calibration mechanism.

Questions for authors

Q1. The proposed framework builds upon established techniques, including latent Gaussian modeling, posterior inference, and negative-control-based nuisance estimation. Could the authors further clarify the key methodological contributions of SceneCal beyond adapting and integrating these existing techniques for multimodal face anti-spoofing? In particular, how does the proposed disjoint calibration mechanism differ fundamentally from existing control-based conditional probabilistic modeling approaches?

Q2. The paper emphasizes that propagating posterior uncertainty is a key component of SceneCal. However, the current ablation studies do not clearly isolate its contribution from that of using disjoint control regions. Could the authors provide a more direct comparison between control-only point estimation and control-only posterior inference, while keeping the control regions and feature representations identical? This would help clarify the independent contribution of posterior uncertainty modeling.

Q3. The framework assumes that non-target control regions are not systematically affected by the presentation attack. However, this assumption may be violated in large-screen replay attacks or scenarios where presentation devices introduce spatially correlated illumination or reflection changes. Unlike isolated control outliers, such coherent changes may not be effectively identified by Huber-based reweighting. Could the authors clarify how SceneCal distinguishes these attack-induced variations from genuine acquisition-state changes? Additional evaluations under structured or spatially correlated control contamination would help assess the robustness and practical applicability of the proposed framework.

Reviewer_91ttOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SceneCal for multimodal face presentation attack detection. Rather than estimating acquisition conditions from the face itself, SceneCal infers a shared latent acquisition state from heterogeneous non-target regions and evaluates facial observations under the resulting predictive distribution. The model is trained on bona-fide samples and incorporates robust control weighting and partial observations. A linear-Gaussian analysis motivates disjoint calibration. Experiments report improved unseen-attack and cross-dataset performance, alongside tests of control availability, diversity, contamination, and acquisition sensitivity.

Critical strengths
  1. The separation between acquisition estimation and target verification is a clear and useful contribution. Section 4 distinguishes plug-in adaptation, target-only marginalization, and control conditioning within the same model, explaining when acquisition-aligned attack evidence can be attenuated. Importantly, the analysis accounts for different null degrees of freedom rather than equating increased noncentrality with improved detection power.

  2. The experiments examine the proposed mechanism beyond aggregate detection accuracy. The control-budget and diversity studies, condition-matched donor replacement, and comparisons with target-plus-context classification help distinguish calibration from simply adding background features. The matched public-benchmark comparisons and the explicit reporting of weaker known-instrument performance also make the empirical assessment more balanced.

Critical weaknesses
  1. The implementation of disjointness is insufficiently specified. Section 3 places controls outside an expanded face guard, but does not identify the feature extractors or their spatial receptive fields. Spatially separate regions do not necessarily yield target-independent measurements if features are computed using surrounding facial content. Since excluding target evidence from calibration is central to the method, the paper should explain how this separation is maintained throughout feature extraction, rather than only during region selection.

  2. The statistical support for the relatively small public-benchmark gains needs clarification. Section 5.3 reports paired video-bootstrap intervals for improvements of 0.38 and 0.67 ACER points, but does not explain how dependencies between recordings of the same subject or session are handled. Three-run standard deviations address training variation, but not this sampling dependence. The missing per-fold and uncertainty appendices prevent assessing whether the reported advantage is consistent across independent evaluation units.

Questions for authors
  1. Which feature extractors produce the modality descriptors and control-type descriptors? Are regions encoded independently, or extracted from full-frame feature maps? Please clarify whether facial content can influence the control measurements through convolutional receptive fields or attention.

  2. What is the resampling unit for the confidence intervals in Section 5.3? Are subjects or sessions grouped during bootstrapping, and how are the leave-one-attack-out folds aggregated? Please provide the corresponding details and per-fold results referenced in the appendices.

  3. Section 3.4 correctly describes the robust covariance as an uncertainty proxy rather than an exact posterior. What do the Appendix G diagnostics show about false-positive calibration across control counts and missing-modality patterns? This matters for assessing whether uncertainty propagation remains reliable beyond the clean Gaussian setting.

Reviewer_edoeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

To address the issue where environmental variations interfere with multimodal relationships in multimodal face anti-spoofing, this paper proposes SCENECAL. This approach leverages heterogeneous background regions surrounding the face to independently infer acquisition conditions. By fixing these conditions before examining facial features, SCENECAL decouples environmental calibration from target verification.

Critical strengths

The author reveals that in face recognition and anti-spoofing tasks, if the environmental changes and the characteristics of an attack align in the feature space, the test-time adaptation mechanism may misinterpret the attack as an environmental shift, thereby causing the detection to fail.

Critical weaknesses
  • The paper is poorly written and difficult to follow. For instance, in Section 5.4, the statement "A target+context supervised head improves Protocol B from 8.91 to 7.54% ACER" lacks an explanation of the specific network architecture. Furthermore, the experimental setup is difficult to comprehend.

  • There is a gap between the paper’s theoretical analysis and the actual system. The practical implementation incorporates a significant amount of non-Gaussian and non-linear processing.

Questions for authors
  • The paper assumes the presentation medium does not systematically alter non-target regions. However, when the medium covers the background, is the acquisition state inferred from controls still reliable?

  • Table 2 shows that pursuing generalization to unseen attacks in Protocol B degrades performance on HQ-MMFAS Protocol A compared to CTNet and MMDG++. This trade-off requires further analysis.

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

Related papers

Discussion (0)

Sign in to comment.