Robust Online Signature Verification Using Long-term Recurrent Convolutional Network

The explosively increasing use of personal computing devices that contain a touchscreen as input interface and the inconvenience of manually pressing password on the devices lead to studies on alternative biometric authentication methods. Among these, handwritten signature has an advantage that using signature as a means of authentication is already familiar to people, but it also has disadvantage that forger can easily attempt to imitate the signature of others. In this context, the purpose of this paper is to boost the forgery verification performance using the machine learning approach. This paper proposes an online signature verification using long-term recurrent convolutional network (LRCN) that ensures extracting distinguishable features between genuine and forged signature. In the proposed method, CNN and time interval embedding are used for feature extraction of signature strokes and LSTM is used for modeling long-term temporal characteristics of stroke sequences. Besides, a forgery-sensitive loss is proposed for robust signature verification against forged signatures. With the loss function, the proposed network provides a feature vector of signature to support vector machine-based classifier for distinguishing a genuine signature from forgeries. The signature verification experiments on the SUSIG dataset show that our proposed work outperforms other state-of-the-art methods on forgery verification performance.