Deep Mamba Siamese Network with Feedforward Layers for Robust Online Signature Verification

Abstract

In the realm of digital biometric authentication techniques, online signature verification hold promise as an alternative to traditional offline methods. But due to the nature of data and intraclass variability, online signature verification remains to a difficult task. Sequential models either lack the ability to fit the data (LSTMs) or are computationally very expensive (Transformers). To address this gap, this paper addresses a deep Siamese Neural Network that users Mamba SSM as the backbone connected to feed forward layers for 1v1 Signature verification. Mamba SSM is a recent development in State Space Model architectures that scales linearly with sequence length while giving performance comparable to transformers. Due to use of the Mamba backbone the model proves to be a fast, lightweight, while still accurate method of online signature verification. Our implementation gave an accuracy 0f 80.5% on a 20% test set of the MCYT100 dataset.