Enhanced-TypeNet for Biometric Keystroke Authentication Using Key Embedding
Mahmoud Bahaa, Fahad Alharbi · Engineering Technology & Applied Science Research · 2025
Deep learning models have demonstrated significant success across various domains, including authentication tasks such as biometric authentication using keystrokes. This paper introduces an advanced model, named Enhanced-TypeNet, based on Long Short-Term Memory (LSTM) with an embedding layer. This model exhibited exceptional performance in large-scale free-text scenarios. Enhanced-TypeNet is an improved version of the original TypeNet model, which had already achieved state-of-the-art results. Model training employed two distinct learning approaches, distinguished by their loss functions: softmax and triplet loss. Building on this foundation, Enhanced-TypeNet shows notable improvements, especially in scenarios with limited enrollment sequences. In the most challenging scenario, involving only one enrollment sequence, the proposed model demonstrates significant improvements in the Equal Error Rate (EER) of 8% and 2% for softmax and triplet loss, respectively, on physical devices compared to the original model. On touchscreen devices, the model achieves even greater enhancements, with EER improvements of approximately 10% and 9% for softmax and triplet loss, respectively. These findings highlight the efficacy of Enhanced-TypeNet across diverse authentication scenarios, emphasizing its potential for real-world applications.