acceptodds
Under review as a conference paper at ICLR 2027

PRIME: PROTOTYPE-ROUTED IMMUTABLE MONOTONE EXPERTS

Abstract

In continual face anti-spoofing (FAS), a router sends each frame to an expert, but bonafide data from later sessions can stop it from sending a known attack to its expert. We propose Prototype-Routed Immutable Monotone Experts (PRIME) to keep these attack routes. When an attack class arrives, PRIME freezes its routing test and the bonafide evidence that the test uses, so later sessions cannot remove a route. We prove this property, conditional zero forgetting on attack routing (ZF-A). The routing test works in a fixed embedding space. PRIME learns this space with Joint Token Attention, a novel readout that weights every token of every layer of a frozen backbone. We pre-train this readout on FFHQ-SynthFAS, our new paired synthetic FAS corpus that we will release. We evaluate PRIME on two protocols. On OCIM leave-one-out with one source per session, PRIME has a mean HTER of 1.45% and outperforms the best continual (10.66%) and joint-training (1.82%) methods. It has the highest (99.97%) and (98.63%), above joint training on the same backbone. On a chronological chain of three public datasets, PRIME has a mean HTER of 2.21% on five unseen datasets and outperforms training-free continual classifiers on the same features. They reverse 1.5% to 15.7% of attack decisions per session, but PRIME does not lose any route.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Face Anti Spoofing, Continual Learning Systems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UQvVOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRIME, a continual learning framework for face anti-spoofing that aims to prevent previously established attack routes from being lost as new sessions arrive. PRIME freezes each attack prototype set together with the bona fide reference used by its routing rule, providing a theoretical guarantee of conditional zero forgetting on attack routing (ZF-A). It also introduces Joint Token Attention (JTA), a learnable readout that aggregates patch and register tokens across all layers of a frozen DINOv3 backbone, pretrained on a newly constructed synthetic dataset, FFHQ-SynthFAS, containing identity-matched attack and bona fide image pairs. Experiments on two protocols show strong cross-domain performance, with a mean HTER of 1.45% on the OCIM leave-one-out benchmark and 2.21% on five unseen datasets in a chronological three-session setting. Comparisons with continual learning baselines also show that PRIME preserves established attack routes without observed reversals on the evaluated samples.

Critical strengths
  • A clearly defined and practically relevant objective. The paper addresses a security-critical issue in continual face anti-spoofing: adding new data should not invalidate previously established attack routes. The formal ZF-A guarantee gives the method a clear objective beyond improving average classification accuracy.

  • Broad empirical evaluation with useful controlled comparisons. The paper evaluates PRIME on two protocols and multiple unseen datasets. In particular, the comparison of routing rules on shared frozen features helps isolate the contribution of the proposed routing mechanism. The paired synthetic dataset also attempts to reduce confounding factors by matching attack and bona fide images in identity and capture conditions.

Critical weaknesses
  • The routing guarantee does not directly guarantee the retention of correct classification decisions. ZF-A ensures that an established expert route cannot be removed, but it does not guarantee that the final prediction remains correct. Indeed, the paper reports that some bona fide frames are routed to attack experts (4.9% in Protocol 1 and 3.3% in Protocol 2). Although the authors acknowledge this limitation, the practical trade-off between route permanence and persistent false attack routing is not sufficiently characterized. More analysis of its impact on final classification errors and deployment reliability would strengthen the central claim.

  • The contribution of the individual methodological components and the fairness of the main comparisons remain insufficiently established. Although Table 4 provides a controlled comparison of routing rules on shared features, several headline comparisons involve methods with different backbones or training data. For example, the Protocol 2 comparison includes published baselines trained on CelebA-Spoof alone, whereas PRIME uses three sequential training datasets. In addition, the analysis does not directly compare JTA with simpler alternatives, such as last-layer token pooling or averaging features across layers. Consequently, it remains difficult to determine how much of the improvement comes from the immutable routing mechanism, the learned projector, and the synthetic pretraining data, respectively.

Questions for authors
  • Can the authors quantify the impact of permanently retained attack routes on final classification errors, particularly for bona fide frames that continue to be routed to attack experts? How does this trade-off evolve as more sessions are added?

  • Can the authors provide controlled ablations of JTA against simpler readout strategies, such as last-layer pooling and cross-layer averaging, to better establish the contribution of the proposed projector?

  • How does PRIME scale to substantially longer continual learning sequences? Since the number of expert evaluations can grow linearly with the number of stored experts in the worst case, it would be useful to report inference cost, memory usage, and routing behavior under a larger number of sessions.

Reviewer_md7pOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a problem in face presentation attack detection (PAD): the incorporation of new data causes previously learned attacks to lose their existing expert routing. PRIME first trains a readout projector on top of a frozen DINOv3 backbone via Joint Token Attention (JTA) on the newly constructed FFHQ-SynthFAS paired synthetic data, after which the feature space is fixed. Each incremental stage establishes an attack prototype set, a bonafide prototype set, and a binary expert. Through append-only bookkeeping, the system guarantees that gates already triggered, together with the earliest responsible expert, never vanish. Predictions from multiple triggered experts are averaged; when no expert is triggered, the router score is used instead. The authors validate the method on the OCIM sequential cross-domain protocol, on a temporally ordered chain composed of three public datasets, and on multiple unseen datasets. The paper reports an average HTER of 1.45% on OCIM, and, under the second protocol, a final average HTER of 2.21% across five unseen datasets.

Critical strengths
  1. Identifies a clear and verifiable failure point of continual learning, and provides an explicit retention mechanism. Freezing the functions on either side of the attack gate decouples the retention of existing routing from how subsequent bona fide data may grow.
  2. Demonstrates promising cross-domain potential. Beyond OCIM, the paper also reports results on multiple unseen datasets, and provides comparisons against a single expert, sequential fine-tuning, and a prototype classifier under the same frozen backbone. Identity-disjoint paired synthetic pretraining is likewise a useful direction for data construction.
Critical weaknesses
  1. The reversed frames metrics in Table 4 are not directly comparable , and the empirical link between routing retention and eventual retention is under-established. According to Appendix G (p. 21, lines 1101–1105), the reversed frames reported for NCM, FeCAM, and RanPAC is defined as "an attack decision turning into a bonafide decision," whereas the reversed frames reported for PRIME is defined as "the loss of an existing Case-1 route." The former is a change in the final label; the latter is a change in an internal state. PRIME's 0% therefore cannot be combined with the baselines' 1.5%–15.72% to support the claim that detection decisions are more stable. Under the definition "an attack decision turning into a bonafide decision," what would PRIME's rate be?
  2. The experiments do not sufficiently isolate the contribution of PRIME's individual components. PRIME combines several potentially powerful components: a large pretrained DINOv3 ViT-H backbone, a learned JTA projector, extensive synthetic FAS pretraining, and the proposed routing mechanism. Yet it remains unclear how much each component contributes to the reported performance.
Questions for authors

1.Verifying the core retention effect: If PRIME were evaluated under the same protocol as NCM, FeCAM, and RanPAC — that is, by measuring the proportion of attacks correctly detected in the previous stage that are classified as bona fide in the next stage — what is PRIME's reversal rate?

2.Contribution analysis of individual components: How much does each component of PRIME contribute to performance? How much of the improvement stems from the immutable routing, and how much from the remaining components?

3.Performance on the widely used, larger-scale WCS datasets (WMCA[3], CASIA-SURF CeFA[2], and CASIA-SURF[1]). PRIME's performance on the OCIM dataset is outstanding and impressive. How does PRIME perform when evaluated on the WCS datasets, and does its advantage still hold? Including this would make the conclusions considerably more convincing.

[1] 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.

[2] 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.

[3] 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_RFcuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PRIME, a continual face anti-spoofing framework that preserves previously established expert routes as new data are introduced. PRIME uses Joint Token Attention to extract features for routing. The authors also introduce FFHQ-SynthFAS, a synthetic dataset used to pretrain the routing projector.

Critical strengths
  1. The method effectively preserves previously established expert routes as new data are introduced.
  2. The experimental results show substantial performance gains.
Critical weaknesses

The method does not guarantee correct predictions for new data routed to existing experts. Highly realistic new attacks may be routed to existing experts and misclassified as bonafide by most of them. Conversely, bonafide samples from new domains may also be routed to existing experts and misclassified as attacks.

Questions for authors
  1. The routing gate for each expert appears to function as a face anti-spoofing classifier in its own right. The overall framework therefore seems similar to an approach that trains one expert per stage and aggregates the predictions of all experts at inference. A direct ablation against this baseline would help clarify the additional benefit of the routing mechanism.
  2. Compared with other methods in Protocol 1, PRIME uses additional information, including the FFHQ-SynthFAS dataset and fine-grained attack-category labels (like “OULU print 1”) from OCIM. Direct comparisons may therefore not be entirely fair.
  3. The main contribution appears to lie more in continual learning than in face anti-spoofing itself. The paper should therefore compare against more continual learning methods and consider evaluation on established continual learning benchmarks.
  4. Missing references

[1] Rehearsal-Free and Efficient Continual Learning for Cross-Domain Face Anti-Spoofing. TPAMI 2025.

[2] Confidence aware learning for reliable face anti-spoofing. TIFS 2025.

[3] Generalized face liveness detection via de-fake face generator. TPAMI 2025.

[4] Detection and Continual Learning of Novel Face Presentation Attacks. ICCV 2021.

Comments on the paper

1 comment
Area_Chair_fvoGOpenReview ↗

Dear Reviewer md7p, thanks for your timely reply. Dear Reviewers RFcu and UQvV, could you please indicate whether the authors' response well addressed all of your suggestions or concerns, and update your final score? Thank you!

Dear ALL,

Dear Reviewer md7p, thanks for your timely reply. Dear Reviewers RFcu and UQvV, could you please indicate whether the authors' response well addressed all of your suggestions or concerns, and update your final score? Thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.