Scholarlink

<p>Dynamic Time Warping (DTW) is a strong baseline for online signature recognition, but naive 1&#xa0;:&#xa0;<i>N</i> identification scales linearly with the number of enrolled identities. In parallel, feature-domain watermarking for provenance and integrity benefits from knowing which descriptors are functionally relied upon by the recognition pipeline. We study an interpretable coarse-to-fine computational architecture, inspired by hierarchical screening principles, on the MCYT online signature corpus (330 users). A fast tree-ensemble preselector produces a Top-<i>K</i> candidate shortlist from a 230-D handcrafted feature vector, and DTW is applied only within the shortlist. Preselectors are benchmarked under repeated resampling of genuine probes, and model reliance is analyzed using interventional SHAP on a claimed-class posterior-like score, complemented by family- and feature-level ablations. The tree-ensemble preselector achieves high shortlist coverage with a compact Top-<i>K</i> list, enabling substantial DTW pruning while preserving identification accuracy after DTW re-ranking. Explainability analyses show that pressure-related descriptors dominate the preselector evidence, and ablations indicate redundancy within the pressure subspace, motivating more selective embedding strategies than broad pressure-domain modification. Explainable Top-<i>K</i> gating is an efficient and analyzable DTW-pruning mechanism for behavioral biometrics. The combined SHAP+ ablation evidence yields actionable watermark-safe guidance: avoid broad embedding in pressure-domain descriptors and prefer embedding locations with low attribution and low ablation-induced performance loss, particularly at low False Acceptance Rate (FAR) under skilled forgeries.</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

Explainable Top-K Gating for DTW-Based Online Signature Identification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月26日

查看全文
  • Marcos Faundez-Zanuy

摘要

Dynamic Time Warping (DTW) is a strong baseline for online signature recognition, but naive 1 : N identification scales linearly with the number of enrolled identities. In parallel, feature-domain watermarking for provenance and integrity benefits from knowing which descriptors are functionally relied upon by the recognition pipeline. We study an interpretable coarse-to-fine computational architecture, inspired by hierarchical screening principles, on the MCYT online signature corpus (330 users). A fast tree-ensemble preselector produces a Top-K candidate shortlist from a 230-D handcrafted feature vector, and DTW is applied only within the shortlist. Preselectors are benchmarked under repeated resampling of genuine probes, and model reliance is analyzed using interventional SHAP on a claimed-class posterior-like score, complemented by family- and feature-level ablations. The tree-ensemble preselector achieves high shortlist coverage with a compact Top-K list, enabling substantial DTW pruning while preserving identification accuracy after DTW re-ranking. Explainability analyses show that pressure-related descriptors dominate the preselector evidence, and ablations indicate redundancy within the pressure subspace, motivating more selective embedding strategies than broad pressure-domain modification. Explainable Top-K gating is an efficient and analyzable DTW-pruning mechanism for behavioral biometrics. The combined SHAP+ ablation evidence yields actionable watermark-safe guidance: avoid broad embedding in pressure-domain descriptors and prefer embedding locations with low attribution and low ablation-induced performance loss, particularly at low False Acceptance Rate (FAR) under skilled forgeries.