Abstract
Iris recognition is widely used due to its high stability and distinctiveness. However, many systems struggle with reliability against presentation attacks, limiting their use in high-security scenarios. Existing state-of-the-art algorithms often rely on iris texture features extracted from raw grayscale space, where differences between real and fake samples are subtle, often restricting detection to single-source spoofs. This paper proposes an iris liveness detection method based on an enhanced gray-level image space. By using a Residual Network (ResNet) to map raw iris images into a separable gray-level space, the discriminative power of features is significantly increased. Features are extracted using a pre-trained LightCNN-4 and the model is trained with a hybrid Triplet and Softmax loss. Experimental results show near-perfect accuracy on closed-set benchmarks and superior generalization in open-set scenarios, outperforming existing methods.
Introduction
While iris biometrics offer exceptional security, they are vulnerable to presentation attacks (PAs) such as printed images, cosmetic contact lenses, or prosthetic eyes. A significant challenge in the field is the "open-set" problem: models trained on specific spoofing types often fail when encountering novel, unseen attacks. This failure stems from the significant overlap of real and fake iris features in the standard raw grayscale domain. This research introduces a spatial mapping technique designed to "magnify" textural discrepancies, making real and fake irises inherently more separable.
Methodology
Enhanced Gray-level Space Generation
The core of the proposed approach is a generator based on the ResNet architecture. Instead of working directly with raw 1D grayscale iris images, the model learns a transformation into an "enhanced" grayscale space. This learned mapping is optimized to highlight discriminative textures—such as the artificial dot patterns in printed irises or the unnatural edges of contact lenses—while preserving the organic complexity of genuine iris tissue.
Feature Extraction and Learning
Feature extraction is performed using a pre-trained LightCNN-4 (Light Convolutional Neural Network). To ensure the generator learns to produce truly useful images, the parameters of the LightCNN-4 are kept fixed during the training of the spatial mapping component.
The training process utilizes a hybrid loss strategy:
- Triplet Loss: This loss function focuses on local sample relationships. It minimizes the distance between "real" iris samples (positives) while maximizing the distance between "real" and "fake" samples (negatives), creating a clear safety margin.
- Softmax Loss: This provides a global classification boundary. Its inclusion stabilizes the training process and accelerates convergence compared to using Triplet loss in isolation.
Experimental Results and Discussion
The method was evaluated using three major datasets: ND-Contact (cosmetic lenses), CRIPAC-Printed-Iris (high-resolution prints), and CASIA-Iris-Fake (multi-source attacks including GAN-generated fakes).
Detection Performance
- Closed-set Accuracy: The model achieved 100% accuracy on the ND-Contact dataset and 99.75% on the CRIPAC dataset.
- Open-set Generalization: In cross-type testing (training on one spoof type and testing on others), the model maintained high performance, reaching 98.94% and 99.06% accuracy. This demonstrates a significant improvement over baseline models like HMC and MT-PAD, which typically show a sharp decline in performance when faced with unknown attack vectors.
Training Efficiency
The hybrid loss approach proved more efficient than Triplet loss alone. In the experimental environment, the inclusion of Softmax loss reduced the average training iteration time from 166.5s to 153.2s.
Conclusion
By transforming iris images into an enhanced gray-level space, this method effectively addresses the difficulty of distinguishing real and fake irises in raw grayscale. The combination of spatial mapping and a hybrid loss function not only improves detection accuracy but also significantly enhances the model's ability to generalize to unknown spoofing types. Future work will focus on expanding dataset diversity to include emerging "deepfake" synthetic eyes and optimizing the architecture for real-time deployment on edge devices.
Summary updated