On Neural Networks for Biometric Authentication Based on Keystroke Dynamics

Nowadays, passwords have become closely associated with our daily activities. However, the development of technology also increases the risk of password leak. For example, the graphics processing unit (GPU)-parallel-computing-based brute force attack and birthday attack algorithms have greatly reduced password security; in addition, passwords are usually transmitted through wired or wireless communication media and thus are vulnerable to attack and easily exposed to illegal users. In this study, we propose a biometric authentication method to identify and block illegal users, even if the entire password is exposed. Our method simultaneously records scan codes and the keystroke sequence of passwords; furthermore, by deep learning of convolutional neural networks (CNNs), it can effectively distinguish legal users from illegal users. We first compare recognition rates between the CNN and the neural network (NN) and prove that the CNN is the better choice. The experimental results show that the proposed CNN model can block all illegal users even if the password is known by them. By using equal amounts of password data from legal and illegal users, the average login failure rate of legal users is 6%, and they can always enter passwords again to be admitted. Finally, by GPU parallel computing, we further accelerate the system performance by 4.45 times.