Abstract
The ubiquitous use of smartphones has spurred the research in mobile iris devices. Due to their convenience, these mobile devices are also utilized in unconstrained outdoor conditions. This scenario has necessitated the development of reliable iris recognition algorithms for such an uncontrolled environment. Additionally, iris presentation attacks such as textured contact lens pose a major challenge to current iris recognition systems. Motivated by these, this paper presents two key contributions. First, a new Unconstrained Multi-sensor Iris Presentation Attack (UnMIPA) database is created. It consists of more than 18,000 iris images of subjects wearing textured contact lens and without wearing contact lenses captured in both indoor and outdoor environment using multiple iris sensors. The second contribution of this paper is a novel algorithm, DensePAD, which utilizes DenseNet based convolutional neural network architecture for iris presentation attack detection. In-depth experimental evaluation of this algorithm reveals its superior performance in detecting iris presentation attack images on multiple databases. The performance of the proposed DensePAD algorithm is also evaluated in real-world scenarios of open-set iris presentation attacks which highlights the challenging nature of detecting iris presentation attack images from unseen distributions.
Introduction
Iris recognition has long been considered a gold standard for biometric authentication due to the unique and stable patterns of the human iris. However, the transition of this technology from high-end, dedicated scanners to ubiquitous mobile devices like smartphones has introduced significant challenges. Chief among these are uncontrolled environments and presentation attacks (PAD).
While mobile iris scanners perform reliably in controlled, indoor settings, their performance often degrades in outdoor environments where high illumination and unpredictable reflections distort iris features. Simultaneously, the use of "textured" or cosmetic contact lenses has emerged as a sophisticated presentation attack. These lenses can be used to obfuscate a user's identity or impersonate another individual by overlaying a synthetic pattern on the natural iris. Most existing PAD algorithms are trained on "clean" datasets captured in laboratory settings, making them "brittle" when faced with real-world variability. This research addresses these gaps by introducing a robust detection framework and a comprehensive multi-sensor dataset.
Proposed Method: DensePAD
To combat the complexities of uncontrolled environments, the authors propose DensePAD, a deep learning architecture based on Densely Connected Convolutional Networks (DenseNet).
The DensePAD framework employs a 22-layer deep structure organized into three "Dense Blocks." Unlike traditional Convolutional Neural Networks (CNNs) where layers only connect to the immediate successor, each layer in a Dense Block receives feature maps from all preceding layers. This dense connectivity offers several advantages:
- Feature Reuse: It encourages the network to reuse features learned in earlier layers, leading to a more efficient representation.
- Parameter Efficiency: It significantly reduces the number of parameters compared to standard CNNs, making the model easier to train and less prone to overfitting.
- Low-Level Detail Preservation: By maintaining "low-level" structural information throughout the network, DensePAD is better equipped to spot the artificial, repetitive patterns found in printed lenses—details that high-level abstract features might otherwise miss.
The model was trained using the Adam optimizer with a growth rate of 12. Input images are normalized to 120 x 160 pixels, and the final layer uses a Sigmoid activation to classify the input as either "Real" or "Attack."
The UnMIPA Database
A core contribution of this work is the creation of the WVU UnMIPA (Unconstrained Multi-sensor Iris Presentation Attack) database. This is currently the largest dataset of its kind, designed specifically to bridge the gap between laboratory success and real-world reliability.
The database contains 18,706 images from 81 subjects. It is unique in its inclusion of images captured across varying lighting conditions—from controlled indoor environments to volatile outdoor settings. Furthermore, the data was collected using three different mobile-grade sensors, ensuring the algorithm's robustness across different hardware platforms. Lux meter measurements confirmed that outdoor intensity values are not only higher but significantly more volatile, presenting a realistic challenge for PAD systems.
Experimental Results
DensePAD was evaluated against traditional methods (such as Local Binary Patterns) and established deep learning models like AlexNet. The results consistently demonstrated DensePAD's superiority:
- Overall Accuracy: On the new WVU UnMIPA database, DensePAD achieved a Total Error of 1.93%, significantly outperforming the next best traditional algorithm, DESIST, which had an error rate of 7.87%.
- Attack Detection: The model reached an APCER (Attack Presentation Classification Error Rate) of 4.02%, meaning it successfully identified nearly 96% of fraudulent attempts even in uncontrolled settings.
- Open-Set Generalization: A critical test for any PAD system is its performance on "unseen" lens brands. In open-set evaluations, DensePAD maintained low error rates (between 2.8% and 3.9%), proving it learns the fundamental "signature" of synthetic lenses rather than just memorizing specific manufacturers.
Conclusion and Future Work
DensePAD represents a significant step forward in making mobile iris recognition secure for the real world. By leveraging dense connectivity, the model achieves state-of-the-art performance in detecting textured contact lenses across diverse sensors and lighting conditions.
However, the "Open-Set" challenge remains partially unresolved. Certain lens colors, particularly Gray, proved harder to detect when not included in the training set (showing an APCER of 10.63%). Future research will focus on evaluating even more diverse deep learning architectures and expanding the variety of sensors to keep pace with the rapidly evolving smartphone market, aiming to further reduce error rates for the most "stealthy" lens types.
Summary updated