Scholarlink

Iris recognition has become a crucial technology in modern security authentication systems. To effectively address the challenges of image anti-spoofing and tampering detection, this paper introduces the MNv4s-ECA (MobileNetV4Small-Efficient Channel Attention) model. The proposed architecture adopts MobileNetV4’s Universal Inverted Bottleneck (UIB) as the backbone, which integrates variants including Inverted Bottleneck (IB), ConvNext-Like, and Extra Depthwise (ExtraDW). An Efficient Channel Attention (ECA) mechanism is incorporated to enhance feature extraction capabilities, selected based on experimental validation. The model is trained using the Adaptive Moment Estimation with Weight Decay (AdamW) optimizer, in conjunction with a composite learning rate scheduling strategy. Experimental results on the IIITD and CASIA-IrisV4 datasets demonstrate that the model achieves a classification accuracy of 99.96% and a perfect authenticity accuracy of 100%, with only 2.5 million parameters. Comparative analyses confirm that the proposed model outperforms existing state-of-the-art networks, maintaining a lightweight structure while achieving high accuracy in detecting forged and tampered iris images.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning-Based Approaches for Iris Image Spoofing Prevention and Tamper Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xiaodong Zhu,
  • Ying Chen,
  • Junkang Deng,
  • Zhijie Chen,
  • Changle He

摘要

Iris recognition has become a crucial technology in modern security authentication systems. To effectively address the challenges of image anti-spoofing and tampering detection, this paper introduces the MNv4s-ECA (MobileNetV4Small-Efficient Channel Attention) model. The proposed architecture adopts MobileNetV4’s Universal Inverted Bottleneck (UIB) as the backbone, which integrates variants including Inverted Bottleneck (IB), ConvNext-Like, and Extra Depthwise (ExtraDW). An Efficient Channel Attention (ECA) mechanism is incorporated to enhance feature extraction capabilities, selected based on experimental validation. The model is trained using the Adaptive Moment Estimation with Weight Decay (AdamW) optimizer, in conjunction with a composite learning rate scheduling strategy. Experimental results on the IIITD and CASIA-IrisV4 datasets demonstrate that the model achieves a classification accuracy of 99.96% and a perfect authenticity accuracy of 100%, with only 2.5 million parameters. Comparative analyses confirm that the proposed model outperforms existing state-of-the-art networks, maintaining a lightweight structure while achieving high accuracy in detecting forged and tampered iris images.