Lightweight CNN with wavelet-attention for fingerprint liveness detection

Abstract

Fingerprint authentication has been extensively employed as a biometric method in diverse applications, including smartphones and embedded systems. Despite its advantages, this technology is susceptible to spoofing attacks using materials such as gelatin, posing a significant security risk. Numerous solutions have been proposed, but deep learning-based approaches often face challenges due to their large model sizes, limiting deployment in resource-constrained environments. To address this issue, we developed a lightweight and efficient fingerprint liveness detection model by integrating wavelet-attention with inverted-bottleneck convolution. The proposed method balances computational efficiency with high accuracy, enabling its practical implementation on low-resource devices.The model was designed with only 874,000 parameters and a memory footprint of 4 MB, representing a significant reduction in size compared to conventional deep learning models. The use of wavelet-attention enhances feature extraction by focusing on multi-scale spatial details crucial for distinguishing live and spoof fingerprints. Extensive experiments were conducted on the LivDet dataset and a custom dataset, encompassing fingerprints captured from multiple sensors.The results demonstrated robust performance, achieving an average classification error (ACE) of 2.27 across various sensors, which is competitive with state-of-the-art methods. Additionally, the model exhibited consistent performance in scenarios with limited computational resources, highlighting its efficiency and scalability.These findings suggest that the proposed approach is a viable solution for enhancing the reliability of fingerprint liveness detection, particularly in applications requiring lightweight and resource-efficient models.