MTCLOSVNET: Multi-Teacher Co-Training And Curriculum Learning For Efficient Online Signature Verification

Abstract

Online Signature Verification (OSV) is a critical component of modern digital security, facilitating real-time user authentication in applications such as financial transactions, identity verification, and secure digital documentation. Al-though deep learning–based OSV systems have demonstrated strong capabilities in feature extraction and forgery detection, their deployment on resource-constrained IoT and edge de-vices remains challenging due to high computational demands. To overcome these challenges, we propose MT-CLOSVNet., a novel Multi-Teacher–Student OSV framework that jointly leverages Curriculum Learning and Adaptive Multi-Teacher Knowledge Distillation (AMTKD). The framework employs two complementary teacher models: (i) a Transformer-based teacher that captures long-range temporal dependencies within signature trajectories, and (ii) a CNN–Transformer hybrid that models fine-grained local stroke dynamics. Knowledge transfer to a compact student network is driven by two key mechanisms: the Mutual Agreement Score (MAS), which distills knowledge only from samples where both teachers yield consistent predictions, and the Adaptive Teacher Weighting Mechanism (ATWM), which dynamically adjusts the student’s reliance on each teacher based on relative training performance. Further, a curriculum learning strategy progressively structures the training process to improve convergence stability and generalization. Extensive evaluations on the MCYT-100, SVC, and SUSIG datasets verify the effectiveness of MTCLOSVNet, achieving state-of-the-art Equal Error Rates (EER) of 12.19%, 6.26%, and 8.54% in the skilled-01 protocol. Remarkably, the lightweight student model contains only 3,266 trainable param-eters—a 98.43% reduction compared to recent SOTA models with 206,277 parameters—while maintaining competitive veri-fication accuracy. These results highlight MTCLOSVNet as a robust, scalable, and deployment-friendly solution for real-time OSV, particularly suited for edge and IoT environments.