Scholarlink

<p>Face recognition systems are widely deployed in authentication and access control, yet remain susceptible to presentation attacks including printed photographs, replay videos, and three-dimensional masks. Face Anti-Spoofing (FAS) has received considerable attention as a countermeasure, but cross-domain generalization—particularly under sensor and environmental variation—remains an unresolved challenge for single-modality approaches. Multimodal methods that incorporate auxiliary signals such as near-infrared (NIR) or depth improve detection robustness, but their dependence on dedicated hardware restricts deployment in cost-sensitive and consumer-grade settings. To address this, we propose <i>ViDiff-Attn</i>, a multimodal FAS framework that eliminates the need for physical NIR sensors by synthesizing NIR representations directly from RGB input. A diffusion-based Spectral Translation Module (STM) performs this RGB-to-NIR mapping, and the resulting synthetic features are combined with RGB representations via a cross-modal attention head that assigns modality-specific importance scores adaptively for each input sample. The full pipeline is trained end-to-end on standard RGB cameras without any auxiliary sensing hardware. We evaluate <i>ViDiff-Attn</i> on four benchmarks—CASIA-SURF, CeFA, WMCA, and PADISI-Face—under both intra-dataset and cross-dataset protocols. <i>ViDiff-Attn-SE</i> achieves an ACER of 1.37% and TPR of 73.28% on Protocol&#xa0;4 in the intra-dataset setting, and an ACER of 2.15% with TPR of 38.73% in cross-dataset evaluation, outperforming all multimodal baselines in comparison. Ablation experiments confirm that both the diffusion synthesis stage and the cross-modal attention fusion are individually necessary—removing either component leads to a measurable drop in performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ViDiff-Attn: Single-Camera Multimodal Face Anti-Spoofing via Diffusion-Based Spectral Synthesis and Cross-Modal Attention Fusion

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Preeti Yadav,
  • Ankit Shukla,
  • Manoj Sharma,
  • Mahesh Kumawat

摘要

Face recognition systems are widely deployed in authentication and access control, yet remain susceptible to presentation attacks including printed photographs, replay videos, and three-dimensional masks. Face Anti-Spoofing (FAS) has received considerable attention as a countermeasure, but cross-domain generalization—particularly under sensor and environmental variation—remains an unresolved challenge for single-modality approaches. Multimodal methods that incorporate auxiliary signals such as near-infrared (NIR) or depth improve detection robustness, but their dependence on dedicated hardware restricts deployment in cost-sensitive and consumer-grade settings. To address this, we propose ViDiff-Attn, a multimodal FAS framework that eliminates the need for physical NIR sensors by synthesizing NIR representations directly from RGB input. A diffusion-based Spectral Translation Module (STM) performs this RGB-to-NIR mapping, and the resulting synthetic features are combined with RGB representations via a cross-modal attention head that assigns modality-specific importance scores adaptively for each input sample. The full pipeline is trained end-to-end on standard RGB cameras without any auxiliary sensing hardware. We evaluate ViDiff-Attn on four benchmarks—CASIA-SURF, CeFA, WMCA, and PADISI-Face—under both intra-dataset and cross-dataset protocols. ViDiff-Attn-SE achieves an ACER of 1.37% and TPR of 73.28% on Protocol 4 in the intra-dataset setting, and an ACER of 2.15% with TPR of 38.73% in cross-dataset evaluation, outperforming all multimodal baselines in comparison. Ablation experiments confirm that both the diffusion synthesis stage and the cross-modal attention fusion are individually necessary—removing either component leads to a measurable drop in performance.