Abstract
Iris recognition is widely used in high-security scenarios due to its stability and distinctiveness. However, imaging differences across devices and racial variations create a significant "domain gap" that hinders the generalization of anti-spoofing models. To address this, we propose the IrisGeneral dataset, a comprehensive benchmark comprising 10 subsets from 7 databases, and the Masked-Mixture of Experts (MMoE) method. MMoE utilizes feature masking and cosine distance loss to force models to learn robust structural features rather than overfitting to sensor-specific artifacts. Experimental results demonstrate that MMoE with a CLIP backbone achieves state-of-the-art performance across all evaluation protocols.
Introduction
Iris recognition systems are vulnerable to presentation attacks, such as printed images or cosmetic contact lenses. While deep learning has improved detection, models often fail in real-world deployments because they overfit to specific training conditions. This "domain gap" stems from hardware variation (different infrared fill lights and sensors) and racial diversity (varying iris textures across ethnicities). Current evaluation protocols are often too narrow to measure a model's true cross-domain robustness. This paper introduces a unified framework to bridge these gaps through a rigorous new testing protocol and a novel architectural approach.
The IrisGeneral Dataset and IAS-CDT Protocol
To evaluate generalization effectively, the authors established the Iris Anti-Spoofing Cross-Domain-Testing (IAS-CDT) protocol using the IrisGeneral dataset. This dataset aggregates over 75,000 samples from 10 subsets, covering 6 different collection devices and multiple racial groups.
The protocol is organized into three hierarchical levels:
- Protocol-1 (Average): Measures general performance across all aggregated data.
- Protocol-2 (Cross-Racial): Evaluates ethnic generalization by training on one race (e.g., Asian) and testing on others.
- Protocol-3 (Cross-Device): Uses a "leave-one-device-out" approach to test robustness against completely unseen camera hardware.
Masked-Mixture of Experts (MMoE)
To solve the generalization problem, the authors propose Masked-MoE (MMoE), built upon a CLIP (Vision Transformer) backbone. While standard Mixture of Experts (MoE) can fit complex data, it is prone to "passive overfitting" where experts become too specialized to training noise.
MMoE introduces two key innovations:
- Mask Mechanism: During training, the system randomly masks input tokens for three out of four experts. One "benchmark" expert receives the full, unmasked image.
- Cosine Distance Loss: The model is penalized if the outputs of the masked experts deviate from the benchmark expert. This forces the masked experts to "fill in the blanks," learning the essential structural features of real vs. fake irises rather than relying on superficial pixel-level artifacts.
Key Results and Insights
The MMoE method achieved State-of-the-Art (SOTA) results across the IAS-CDT protocols, significantly outperforming standard models like ResNet50 and vanilla ViT:
- Protocol-1: Achieved a 1.23% ACER (Average Classification Error Rate), a 13% improvement over the baseline CLIP model.
- Protocol-2 (Cross-Racial): Achieved 7.59% ACER, demonstrating superior ethnic generalization compared to ResNet50 (12.38%).
- Protocol-3 (Cross-Device): Maintained a low 4.85% average ACER, proving high robustness when encountering brand-new camera sensors.
Conclusion
The proposed framework provides a much-needed benchmark for iris anti-spoofing generalization. By combining a diverse dataset with the MMoE regularization technique, the model learns to identify the core essence of iris textures. Future work will focus on optimizing the computational efficiency of MoE blocks for edge device deployment and exploring the optimal balance for masking rates to maintain high-quality feature extraction.
Summary updated