A lightweight privacy-preserving iris recognition method
Haorui Wu, Langwen Zhang, Wei Chau Xie · 2025
Recently, biometric identification has gained widespread application, among which iris recognition posses high security and accuracy due to its difficulty of acquisition and high information density. However, large-scale convolutional neural networks struggle to perform real-time computations on low-computational-power platforms. Additionally, while there has been significant research on privacy-preserving face recognition, studies focusing on iris recognition remain scarce. To address this, this paper proposes a lightweight privacy-preserving iris recognition method. The approach involves applying Discrete Cosine Transform (DCT) to iris images, removing low-frequency components, and performing channel shuffle. Furthermore, the lightweight network ShuffleNet V2 is improved by adding ECA layer and MBConv module. The method is trained and tested on the CASIA-Thousand near-infrared iris dataset, achieving an open-set Top-1 recognition accuracy of 98.54% while maintaining model inference speed.