acceptodds
Under review as a conference paper at ICLR 2027

VADAPT: Visual Anchored and Dual-Prompt Tuning for Generalizable Face Attack Detection

Abstract

Generalizable face attack detection must handle not only conventional physical presentation attacks but also rapidly evolving digital manipulations. While pretrained vision-language models (VLMs) offer transferable representations, their direct adaptation to this task remains limited by heterogeneous real/fake distributions and concentrated reliance on a narrow set of visual evidence. To address these limitations, we propose VADAPT, a Visual Anchored and Dual-Prompt Tuning framework for generalizable face attack detection. VADAPT introduces Visual Soft-Anchoring, which uses indexed textual anchors to softly differentiate multiple learnable visual part tokens, encouraging diverse fine-grained evidence exploration without predefined anatomical partitions. In parallel, Dual-Branch Latent Context learns multiple continuous contexts to capture heterogeneous genuine and attack-related representations, while Asymmetric Feature Aggregation adaptively combines their responses using branch-specific aggregation rules. Extensive experiments on UniAttackData and HydraFake demonstrate strong cross-attack generalization. VADAPT achieves state-of-the-art performance on UniAttackData under the official cross-form protocol, attaining 13.45% ACER, compared with 15.13% for the strongest previous baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
face anti-spoofing, face attack detection, domain generalization, prompt tuning, CLIP, vision-language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_JgUyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VADAPT, a vision-language adaptation framework for generalizable face attack detection. It combines Visual Soft-Anchoring (VSA), which aligns learnable visual part tokens with indexed textual anchors; Dual-Branch Latent Context (DLC), which learns multiple prompts for genuine and attack samples; and Asymmetric Feature Aggregation (AFA), which applies different pooling strategies to the two branches. Experiments on UniAttackData and HydraFake show promising cross-attack generalization, with an ACER of 13.45% on UniAttackData compared with 15.13% for the strongest reported baseline.

Critical strengths
  1. The paper addresses a practically important problem: generalizing across physical presentation attacks and digital face manipulations under distribution shifts.
  2. The framework is coherent, and ablations investigate the contributions of textual anchoring, multiple latent contexts, asymmetric pooling, and part-level aggregation.
  3. VADAPT achieves promising results on UniAttackData and strong cross-model and cross-domain performance on HydraFake, using a relatively compact vision-language backbone.
Critical weaknesses
  1. The novelty over existing VLM-based prompt-learning methods is not sufficiently established. The paper positions VADAPT as a new prompt-based framework for generalizable face attack detection, but several closely related VLM/prompt-learning approaches, such as CFPL-FAS, CrossGAP, and recent hierarchical prompt tuning methods, are discussed in the related work but are not included in the main UniAttackData comparison. Instead, Table 1 mainly compares against conventional FAS/FAD methods and MoAE-CR. This makes it difficult to determine whether the gains come from a genuinely new prompting mechanism or simply from applying multi-context and visual prompt tuning to a strong CLIP/FaRL backbone. Direct comparisons with the closest prompt-learning baselines under the same backbone, training data, and evaluation protocol are necessary to substantiate the methodological novelty.

  2. The experiments do not demonstrate that the benefit of Visual Soft-Anchoring specifically comes from textual anchoring. The proposed textual anchors are intentionally generic and differ only by an index, e.g., “a photo of a human face part i”; they are not associated with any semantic facial region. Therefore, the anchoring loss effectively forces different visual tokens toward different fixed directions in the CLIP embedding space, but it is unclear why these directions need to come from language. The current controls use random textual sentences or the same textual anchor for all tokens. These baselines do not isolate the role of the language space from generic representation separation. Stronger controls should include fixed random vectors, orthogonal non-text prototypes, learnable anchor embeddings, or direct pairwise diversity/orthogonality regularization. Without such experiments, the proposed mechanism may be better interpreted as token diversification rather than a specifically vision-language anchoring contribution.

  3. The claim that Dual-Branch Latent Context captures heterogeneous real/fake modes is not sufficiently supported. DLC introduces multiple learnable real and fake contexts, but these contexts are optimized only through the final classification objective, without any explicit diversity, assignment, or orthogonality constraint. Consequently, there is no guarantee that different contexts actually correspond to different latent modes rather than simply providing additional parameters or an ensemble-like effect. The current evidence mainly consists of a performance comparison against a single-context variant and qualitative visualization showing non-identical responses. This is insufficient to establish the claimed latent-mode modeling mechanism. The paper should quantitatively measure context similarity and utilization, analyze whether different attack types or real domains activate different contexts, and compare against parameter-matched multi-prototype or multi-head baselines.

  4. The motivation for the asymmetric Max/Mean aggregation is heuristic and does not logically follow from the presented feature visualization. The paper argues that separated real clusters motivate Max pooling over real contexts, whereas overlapping fake clusters motivate Average pooling over fake contexts. However, the observed geometry in a t-SNE visualization does not directly imply that Max and Average are the appropriate statistical aggregation operators. For example, cluster separation alone does not establish that matching one real prototype should be sufficient, nor does fake-class overlap imply that averaging all fake prototypes is optimal. Although the Max/Avg combination performs best among the four Max/Mean combinations, this only provides empirical model selection rather than validating the proposed reasoning. Comparisons with learnable attention pooling, log-sum-exp pooling, mixture weighting, or other adaptive aggregation mechanisms would be needed to establish that the asymmetric rule reflects an underlying property of the data rather than a dataset-specific heuristic.

  5. The proposed fine-grained evidence exploration mechanism lacks quantitative validation. Figure 3 qualitatively shows diverse Grad-CAM responses, and Table 4 supports the usefulness of indexed anchors. However, distinct attention maps do not necessarily imply complementary forensic representations. Since the anchors differ mainly by an index in a generic text template, it also remains unclear whether their semantic content is essential or whether they simply provide token-specific reference vectors. Quantitative analysis of inter-token diversity and comparisons with alternative or permuted anchors would strengthen the central claim.

  6. The claimed state-of-the-art improvement on UniAttackData is small relative to the reported variability. The headline result improves ACER from 15.13% for MoAE-CR to 13.45% for VADAPT, i.e., an absolute improvement of only 1.68 points. However, the reported deviations are 12.10 for MoAE-CR and 2.51 for VADAPT, which are much larger than the difference between the two mean values. Therefore, the current results do not clearly establish that VADAPT is statistically better than the strongest baseline. The paper should report paired runs or confidence intervals, statistical significance tests, and preferably the physical-to-digital and digital-to-physical results separately. Otherwise, the strong “state-of-the-art” conclusion appears stronger than what the reported statistics support.

  7. The reliability and reproducibility of the improvements need clarification. On UniAttackData, ACER improves from 15.13% to 13.45%, while ACC increases only from 85.41% to 86.02% and EER decreases from 13.81% to 13.53%. Given the reported variability, the statistical reliability of these gains is unclear. The paper does not adequately specify the number of independent runs or the source of the reported standard deviations. Important training details, including the number of epochs, data augmentation, learning-rate scheduling, and threshold selection, are also missing. More complete implementation details and uncertainty estimates are needed.

Minor Issues

  1. Please clarify whether features in the anchoring loss are normalized and whether temperature scaling is applied.
  2. Specify which parameters are frozen or trainable, how visual tokens are inserted into the pretrained ViT, and how positional embeddings are handled.
  3. Report the number of independent runs, define the uncertainty estimates, and clarify the HydraFake aggregation protocol.
Questions for authors

Please refer to Weaknesses.

VADAPT addresses an important problem and presents a coherent framework with promising empirical results and useful ablations. However, the methodological novelty relative to existing prompt-tuning approaches remains insufficiently demonstrated, and the central claim that textual anchoring enables complementary fine-grained forensic evidence is supported mainly by qualitative analysis. The robustness and reproducibility of the reported improvements also require clarification.

For the rebuttal, I encourage the authors to provide matched comparisons with standard prompt-tuning baselines, stronger quantitative evidence for visual-token diversity, and complete details on experimental variability and metric aggregation. These issues do not invalidate the approach, but the current evidence is insufficient to establish its technical contribution and generalization claims fully.

Reviewer_19DHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents VADAPT, a VLM-based framework for detecting face attacks under distribution shift. The authors argue that prior works are limited by heterogeneous real/fake distributions and concentrated reliance on a narrow set of visual evidence. VADAPT has three components: Visual Soft-Anchoring uses indexed textual anchors to guide multiple learnable visual part tokens; Dual-Branch Latent Context learns multiple real and fake prompts without predefined attack categories; and Asymmetric Feature Aggregation combines their responses using different rules for the two branches. The authors evaluate VADAPT on UniAttackData and HydraFake and report strong cross-form and cross-distribution performance.

Critical strengths
  • Unified face attack detection is an important problem because a detector may encounter both physical presentation attacks and digital forgeries. The paper evaluates transfer between these two attack forms, which is a relevant test of generalization.

  • The paper proposes a concrete way to model diverse visual cues and real/fake patterns without assigning prompts to predefined facial regions or attack types. Its ablation results provide some evidence that the visual anchors, multiple contexts, and aggregation choices contribute to the reported performance.

Critical weaknesses
  • My main concern is that the introduction section does not clearly establish the paper’s novelty relative to the most relevant prior work. It motivates the method by arguing that “directly adapting a global CLIP representation and a small set of fixed prompts leaves two task-specific issues unresolved.” However, both the paper’s related-work discussion and my review of relevant studies indicate that several closely related methods already go beyond global representations or fixed prompts. Although the Introduction describes two challenges and outlines VADAPT’s responses, it does not adequately explain which limitations remain in those prior methods, how VADAPT addresses them, or what distinct contribution it makes beyond them.

  • The second concern is that experimental comparisons do not adequately support the claimed advance. Although the Introduction cites HiPTune as Liu et al. (2025), Table 1 does not compare against it, while the paper describes 15.13% ACER as the strongest prior result. HiPTune reports 10.38% ACER on UniAttackData Protocol 2. The authors should clarify whether the protocols and supervision requirements are comparable, qualify the state-of-the-art claim accordingly, and provide comparisons that establish VADAPT’s incremental contribution.

Questions for authors
  • Are HiPTune’s 10.38% ACER on UniAttackData Protocol 2 and VADAPT’s 13.45% ACER in Table 1 directly comparable? Please specify any differences in data splits, attack-type supervision, and metric aggregation. If they are comparable, how should the state-of-the-art claim be interpreted?
  • Could the authors clarify in the Introduction how VADAPT differs from closely related methods that already use adaptive prompts or local visual features, such as FA³-CLIP[1] and HiPTune? Which limitation of these methods specifically motivates VADAPT?

[1] FA3-CLIP: Frequency-Aware Cues Fusion and Attack-Agnostic Prompt Learning for Unified Face Attack Detection

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.