Iris Liveness Detection for Video Replay Attack Using Quality Metrics and EfficientNetB7

Abstract

Translate Article

The replay attack is a potentially serious threat to the dependability of Iris recognition systems, in which authentic biometric characteristics are reproduced from pre-recorded video recordings. To this end, we present a robust Iris Liveness detection framework based on best-frame selection, deep learning classification, and quality metric evaluation. To evaluate the performance, a real dataset of 115 eye samples of the users were used along with fake data. Video frames were presented to Laplacian sharpness variance algorithm for best frame selection. The best frames were categorized by passing them through a fine-tuned EfficientNetB7 model. Furthermore, biometric quality metrics (pupil-Iris ratio, grayscale utilization etc.) were combined to improve the Liveness. As testing dataset is divided into two subset 40 videos and 60 videos. The experiments demonstrated the accuracy of 92.50 % and 91.67% for 40 videos and 60 videos respectively. In addition, EfficientNetB7 exhibited excellent accuracy and robustness in distinguishing between real and spoofed samples. These findings validate our design's generalization across genuine and attack domains, and demonstrate that it can be used as a scalable solution for real-time Iris spoof detection