Lightweight And Low-Latency Inference Based Online Signature Verification

Abstract

An online signature can be viewed as a set of multivariate time-series signals, and it remains a widely used biometric for real-time user verification on memory-constrained devices such as digital signature pads. Recent advances in deep learning (DL) have encouraged its adoption for Online Signature Verification (OSV); however, most DL-based OSV solutions still inherit major deployment challenges: (i) heavyweight models with lakhs of trainable parameters, (ii) limited feasibility for resource-constrained edge devices, and (iii) high inference latency in real-time usage. To address these limitations while retaining the benefits of DL, this work leverages TensorRT, a parallel programming model built on NVIDIA CUDA, to perform post-training optimization of DL models and to accelerate inference. We experimentally analyze the above OSV bottlenecks and then present a superior framework that achieves state-of-the-art (SOTA) performance on two widely used datasets, MCYT-100 and SVC. In particular, the proposed model reports an equal error rate (EER) of 11.98% and 5.03% in the Skilled_01 category for the MCYT-100 and SVC datasets, respectively. The results further confirm that the proposed approach produces lightweight models with reduced inference latency: a 74.76% reduction in model size leads to only a 2.22% reduction in accuracy in the Skilled_01 category of MCYT. Overall, the experimental findings show that the proposed framework yields optimized OSV models that are lightweight, low-latency, and accurate, making them suitable for deployment on resource-constrained edge devices.