acceptodds
Withdrawn from ICLR 2027

Physical Semantic Graph for Face Anti-Spoofing via Cross-Attribute Consistency Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent RGB face anti-spoofing (FAS) methods have achieved remarkable progress by learning discriminative representations from various visual cues. However, most existing approaches remain predominantly feature-driven, focusing on individual cues or their fusion while paying limited attention to the relational structure among heterogeneous physical evidence. In this paper, we introduce a new perspective that characterizes facial liveness through cross-attribute consistency among complementary physical semantics. Specifically, we propose a fixed Physical Semantic Graph (PSG), where each node represents observable or estimable semantic evidence associated with a physical property of facial liveness, including texture, frequency, geometry, reflectance, and physiology-related evidence, while graph edges encode prior relational hypotheses derived from physical dependencies and complementary observations. Based on PSG, we develop a cross-attribute reasoning mechanism that propagates complementary evidence along these predefined relations to learn a relational representation of facial liveness. Furthermore, we introduce Pairwise Physical Consistency Regularization (PPCR), which explicitly encourages the relational patterns of genuine faces to be more consistent than those of presentation attacks. Extensive cross-dataset experiments on multiple public FAS benchmarks demonstrate competitive generalization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Graph learning, physical semantic graph, face anti-spoofing

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_nPJGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The method formulates RGB face anti-spoofing as relational reasoning over complementary physical cues. It builds a Physical Semantic Graph over texture, frequency, geometry, reflectance, and physiological evidence, and uses consistency regularization to distinguish genuine faces from attacks. Cross-dataset experiments show competitive generalization.

Critical strengths

A key strength is the clear conceptual shift from feature-driven fusion to relational reasoning over complementary physical semantics. The Physical Semantic Graph provides an interpretable way to model dependencies among texture, frequency, geometry, reflectance, and physiological cues, which gives the method a stronger physical motivation than conventional cue aggregation. A second strength is the cross-dataset evaluation, which shows that the proposed consistency-based reasoning achieves competitive generalization across multiple FAS benchmarks; this is particularly important because cross-domain robustness is a central challenge in FAS.

Critical weaknesses

However, several issues are required to be clarified:

  1. My major concern is the technical novelty of the proposed method. GCN-based relational modeling has been widely explored, while the current design appears to introduce only limited architectural innovation. The authors should provide a more insightful explanation of the methodological novelty and clarify what fundamentally distinguishes the proposed graph reasoning mechanism from existing GCN-based approaches.

  2. Why are these five physical semantics—texture, frequency, depth, reflectance, and physiology-related cues—selected as graph nodes? The authors should provide stronger justification for this choice and discuss whether these semantics are sufficiently comprehensive to characterize facial liveness or whether other relevant physical cues may have been overlooked.

  3. Only one method published in 2026 is included in the comparison. The authors are encouraged to compare with more recent approaches to better demonstrate the competitiveness and superiority of the proposed model against the latest advances.

  4. Qualitative results should also be provided to illustrate the relative importance of different physical semantics across representative cases and to show how each semantic cue contributes to the final liveness prediction.

Questions for authors

Please clarify several points that could materially affect my assessment: first, what is the key technical novelty of the proposed graph reasoning mechanism beyond existing GCN-based relational modeling, and which specific design choices fundamentally distinguish it from prior graph-based FAS methods; second, what theoretical or empirical basis supports selecting texture, frequency, depth, reflectance, and physiology-related semantics as the five graph nodes, and whether these cues are intended to comprehensively characterize facial liveness; and third, given that only one 2026 method is included, how should the reported performance be interpreted relative to more recent FAS approaches and whether the main conclusions are expected to remain competitive. As a suggestion for future work, qualitative case-level analyses showing how different physical semantics contribute to individual predictions would further improve interpretability.

Reviewer_bzEJOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies cross-domain RGB face presentation attack detection, and proposes characterizing liveness through the relationships among multiple cues rather than performing feature fusion alone. The method extracts five types of descriptors: LBP texture, FFT spectrum, relative depth from a frozen MiDaS-Small, reflectance estimation from a frozen PIE-Net, and normalized chromaticity statistics of skin regions. These form a Physical Semantic Graph (PSG) with five nodes and six predefined undirected edges. After a two-layer GCN updates the nodes, the method builds edge relation representations via feature concatenation, difference, and element-wise product, and then classifies using pooled node and edge representations. PPCR employs a cosine triplet margin loss so that the same edge representation is more similar for bona fide–bona fide pairs than for bona fide–attack pairs. Validation includes four leave-one-domain-out experiments on MCIO, together with topology, node, and component ablations on I&C&M to O, and an analysis of relation similarity distributions on SiW. The paper reports HTERs of 5.6%, 7.1%, 5.2%, and 7.3% on the four target domains.

Critical strengths
  1. The organization of heterogeneous evidence is explicit. All five inputs and all six connections are listed explicitly, and the fixed topology does not need to be constructed from target-domain features. This makes the method's inductive bias inspectable, replaceable, and ablatable, and it also facilitates the study of complementarity among different cues.
  2. Targeted preliminary ablations are provided. Table 2 includes edge-count-matched Random/KNN graphs and a semantic-shuffling control that preserves the connectivity structure; Table 4 uses the same base descriptors and reports that HTER drops from 12.6% to 9.1%, 8.2%, and 7.3% as PSG, relation modeling, and PPCR are successively added.
Critical weaknesses
  1. Table 1 baselines do not match the original papers, so the comparison is unreliable. Many competing methods deviate substantially from their reported results — e.g., in M/C/I/O order, CFPL[1] shows 5.3/8.6/8.4/8.5 here versus 3.09/2.56/5.43/3.33 in the original paper (mean 3.60%, same leave-one-domain-out protocol, no CelebA-Spoof pretraining), against PSG's 6.30% average. This materially changes the assessment of relative competitiveness. If these rows are the authors' own reproductions, they should be labeled as such and the discrepancies explained (several of these methods release no code — how is reproduction correctness established?); if quoted directly, they must be corrected and the conclusions re-evaluated.
  2. The "physical relations" are not disentangled from generic fusion and triplet learning. Eq. (23) is essentially a class triplet constraint anchored on bona fide samples, and edge representations that merely encode the bona fide/attack label may already satisfy it without recovering any physical dependency. The G–G/G–S separation in Fig. 2 is precisely PPCR's optimization target, so it cannot by itself validate the mechanism. The fusion baseline only averages the five embeddings, lacking a capacity-matched concatenation MLP, generic attention fusion, and a same-loss control on non-graph features; Tables 2–4 also cover a single target domain. The gain may thus stem from fusion capacity and generic discriminative regularization rather than the claimed physical relations.

[1] Ajian Liu, Shuai Xue, Jianwen Gan, Jun Wan, Yanyan Liang, Jiankang Deng, Sergio Escalera, and Zhen Lei. Cfpl-fas: Class free prompt learning for generalizable face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 222-232, 2024.

Questions for authors

1.Clarification of baseline provenance: Are the rows in Table 1 taken from the original papers, quoted from other studies, or uniformly reproduced? Please provide the source or the reproduction configuration for each entry, and explain the discrepancies between each competing method and its original reported results. Do the paper's conclusions about competitiveness still hold once the protocols are corrected or aligned?

2.Validation of the mechanism: How much gain remains for PSG when a concatenation MLP or an attention-based fusion, built on the same five descriptors with a comparable parameter count, is trained with the same triplet loss applied to its features? Without PPCR, does the fixed physical topology still outperform the Random, KNN, and Shuffled graphs, and does this advantage reproduce under other leave-one-domain-out splits?

3.Evaluation is limited to MCIO. Results are reported only under the MCIO leave-one-out protocol; adding CASIA-SURF[2], CASIA-SURF CeFA[3], and WMCA[4] would considerably strengthen the evidence.

[2] Shifeng Zhang, Ajian Liu, Jun Wan, Yanyan Liang, Guodong Guo, Sergio Escalera, Hugo Jair Escalante, and Stan Z Li. Casia-surf: A large-scale multi-modal benchmark for face anti-spoofing. TBMIO, 2(2):182–193, 2020.

[3] Ajian Liu, Zichang Tan, Jun Wan, Sergio Escalera, Guodong Guo, and Stan Z Li. Casia-surf cefa: A benchmark for multimodal cross-ethnicity face anti-spoofing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1179–1187, 2021.

[4] Anjith George, Zohreh Mostaani, David Geissenbuhler, Olegs Nikisins, Andre ́ Anjos, and S ́ebastien Marcel. Biometric face presentation attack detection with multi-channel convolutional neural network. TIFS, 2019.

Reviewer_76BkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

he paper proposes a fixed "Physical Semantic Graph" (PSG) for cross-domain face anti-spoofing, in which five semantic cues (texture/LBP, frequency/FFT, MiDaS depth, PIE-Net reflectance, and skin chrominance) are treated as graph nodes with a hand-designed six-edge topology. A two-layer GCN performs "cross-attribute reasoning," and a Pairwise Physical Consistency Regularization (PPCR) term encourages genuine faces to yield more stable edge relations. The central claim is that the gains arise from physically motivated relational structure rather than from feature fusion or generic graph connectivity.

Critical strengths

The motivation — that attacks disrupt the joint configuration of physical cues even when they preserve individual cues — is interesting and well-motivated. The idea of encoding interpretable priors as a sparse fixed topology is appealing, and the evaluation follows the standard MCIO leave-one-out protocol with multiple public benchmarks.

Critical weaknesses

Major Concern 1 — The paper does not establish its central claim. The headline contribution is that gains come from relational graph structure, not from fusing multiple cues. However, the experimental design cannot support this. In Table 4, the "Baseline" is a deliberately weak fusion: the five embeddings are averaged. Moving to "+PSG" (HTER 12.6→9.1) therefore conflates two changes at once — (a) replacing averaging with a learnable, deeper GCN-based fusion, and (b) imposing the physics-derived mask. There is no comparison against strong, fusion-only baselines on the same five features (e.g., concatenation + MLP, bilinear pooling, or cross-attention). Without such an ablation, the observed improvement is at least as consistent with "a better fusion module over complementary cues" as with "relational reasoning," and the paper's key conclusion is unproven.

Major Concern 2 — Evidence actually points toward fusion, not the graph. Table 3 shows that removing any single node degrades performance. This demonstrates that the benefit scales with the number of complementary cues — i.e., it is direct evidence for the fusion hypothesis, which the authors should have isolated but instead cite as support for relational modeling. Likewise, the topology analysis (Table 2) is confounded: PSG+Shuffled and edge-matched Random graphs connect semantically unrelated nodes, so degradation is expected regardless of whether physics-derived relations add real value. A fair test requires a learned adjacency and an attention-fusion variant with matched parameters, both absent here.

Major Concern 3 — Insufficient experimental design. (i) With only five nodes and a two-layer GCN, "cross-attribute reasoning" is essentially sparse, masked message-passing; its incremental value over learnable fusion is minimal, yet this is never measured. (ii) Differences are small and reported without variance or significance tests despite "three seeds" (e.g., AUC gaps of ~1% in Table 1), so it is unclear whether the margins over prior work exceed noise. (iii) Table 1, which substantiates the "competitive/superior" claim, appears incomplete and cannot be verified as presented. (iv) The edge inclusions/exclusions are justified only heuristically; ablations over these design choices are missing, and the PPCR analysis on SiW (Fig. 2) is largely tautological, since PPCR is trained to maximize exactly the reported G–G vs. G–S separation.

Questions for authors

The idea is promising, but the paper's defining contribution — that the graph, rather than multi-cue fusion, drives performance — is neither tested against appropriate fusion baselines nor isolated from confounds. The authors must add strong fusion-only comparisons on identical features, a learned-graph/attention baseline, variance and significance reporting, and a complete comparison table before the relational claim can be credited.

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

Related papers

Discussion (0)

Sign in to comment.