A2002
Title: Explainable writer verification through dynamic handwriting feature analysis
Authors: Seungjin Baek - Pusan National University (Korea, South) [presenting]
Abstract: Writer verification is a crucial task in biometric authentication and forensic handwriting examination, where reproducibility and reliability of decisions are required. While offline handwriting analysis mainly relies on the final written form, online handwriting data provide dynamic information, such as coordinates, pressure, velocity, and acceleration, that reflects individual writing habits and motor patterns. Existing studies have improved verification performance using complex deep learning models; however, the lack of transparency in these models makes their decisions difficult for forensic experts to trust in practical settings. Therefore, writer verification must balance predictive performance with explainability. An explainable writer verification framework is developed using multivariate online handwriting data from the AI Hub online handwriting dataset. A bidirectional LSTM-based model is employed to learn complex dynamic patterns from continuous coordinate, pressure, and kinematic features. To support interpretability, Shapley value-based feature attribution is applied to conduct feature-level interpretation. This analysis shows how each dynamic component contributes to the final verification decision, highlighting specific handwriting movements that support genuine or forged classification. Ultimately, the proposed approach goes beyond performance-oriented prediction and supports transparent forensic decision-making.
