Deepfake In-Air Signature Verification via Two-Channel Model

Hao Yu, Hongxi Wei, Yiming Wang

Abstract: The in-air signature verification system has garnered attention due to its flexibility, friendliness, security, high efficiency, remote accessibility, and contactless usage patterns. However, with the advancement of machine learning, computers can now learn and imitate many things through artificial intelligence. This technology, referred to as “deepfake”, has been used to create fake signatures, raising concerns about the security of the in-air signature verification system. Consequently, research on in-air signature verification technology against deepfakes has become an urgent need. The current challenges are: (1) Most research on in-air signature verification at this stage focuses on human forgery, with a lack of high-performance verification systems for deepfake in-air signatures. (2) Due to the long sequence and small amount of data, there is a lack of good generation methods for in-air signature generation tasks. To address these challenges, this paper proposes an improved in-air feature data model based on the GLNLSTM autoencoder. The signature samples generated by this model are more authentic than the baseline model. Additionally, we introduce an in-air signature verification model based on a two-channel model. The signature semantic feature extraction module of this model uses one-dimensional CNN and bidirectional LSTM to extract dynamic time features. This model achieves the best results in both deepfake verification and artificial forgery signature verification tasks on the SCUT-MMSIG-AIR database.