Attention-Driven CNN-LSTM Hybrid Models for Secure Dynamic Signature Verification

Dynamic signature verification is a crucial biometric verification process, yet traditional methods are not very efficient in forgery detection and signature variation as they focus only on static signature features. In this paper, two hybrid deep learning models CNN-LSTM and CoAtNet are implemented to improve verification accuracy based on a signature's dynamic features such as spatial co-ordinates, pen pressure, velocity, acceleration etc. The hybrid CNN-LSTM works with both spatial and temporal patterns to extract meaningful patterns from dynamic data, while CoAtNet uses convolutional layers and transformer-based layers to perform stronger feature extraction using attention-based mechanisms. Two real world datasets, SIGNDUMP and SignatureFeatures dataset were used for experimentation and analysis. The results indicate that CNNLSTM is 99.22% accurate with big data and outperforms CoAtNet (97.90%). On the contrary, CoAtNet outperforms CNN-LSTM (94.89%) for small datasets and achieves 97.34% accuracy with SIGNDUMP. The findings exhibit the complementary abilities of the two architectures in signature verification. As a result, the proposed hybrid deep learning models enhance forgery detection and security of authentication significantly and are hence suitable for banking, legal verification, and government authentication systems. Future work may consider ensemble neural networks combining these approaches and other attention-based mechanisms for further performance enhancement.