Introduction
Iris recognition systems offer high security but are increasingly targeted by sophisticated Presentation Attacks (PAs). These attacks range from high-resolution printed images and digital displays to cosmetic contact lenses and artificial eyes. Traditional Presentation Attack Detection (PAD) systems typically employ binary classification—training models to distinguish between "real" and "fake" samples. However, this approach often fails to generalize to "unseen" attacks: novel spoofing methods not present in the training data.
To address this, we propose RD-PAD, a novel framework that treats iris spoofing detection as a one-class anomaly detection problem. By focusing on learning the intricate details of bonafide iris images, the system can identify any deviation as a potential attack, regardless of whether that specific attack type was previously encountered.
Methodology: The Relativistic Advantage
The core of RD-PAD is the Relativistic Average Generative Adversarial Network (RaSGAN). Unlike standard GANs, where the discriminator simply classifies images as real or fake, a relativistic discriminator estimates the probability that a real image is "more real" than a fake one. This creates a more stable training process and forces the model to capture finer anatomical details.
One-Class Training (Method-I)
In our primary approach, the RaSGAN is trained exclusively on bonafide iris images. The generator attempts to create synthetic irises, while the Relativistic Discriminator (RD) learns to distinguish real samples from these high-quality synthetic adversaries. This internal competition forces the discriminator to establish an extremely "tight boundary" around the distribution of real iris data.
Fine-Tuning (Method-II)
To further refine the decision boundary and improve cross-sensor performance, we also explored a fine-tuning stage where the discriminator is exposed to a small number of known PA samples. This helps the model adjust its "tight boundary" to better accommodate variations in image capture conditions.
Experimental Evaluation
We evaluated RD-PAD across six diverse datasets: MSU-Iris-PA01, Casia-Iris-Fake, Berc-Iris-Fake, NDCLD15, LivDet15, and LivDet17. These datasets include a wide array of PAs, such as Kindle displays, printed irises, and various cosmetic lenses. All images were pre-processed for alignment and resized to 256x256 pixels.
Key Results and Insights
RD-PAD demonstrated superior generalization compared to traditional binary classifiers like BSIF+SVM and VGG-16.
- Generalization to Unseen Attacks: In scenarios involving unseen cosmetic lenses and Kindle-based attacks, RD-PAD (Method-II) achieved an APCER (Attack Presentation Classification Error Rate) of 21.79% at a 5% BPCER. In comparison, the baseline BSIF+SVM method reached an APCER of over 91%, failing to detect the novel attacks.
- The Generator as a Synthetic Adversary: The GAN's generator acts as an infinite source of "near-miss" fakes, allowing the discriminator to learn the edges of the bonafide distribution without requiring physical spoof samples for every possible attack type.
- Future-Proofing: By framing PAD as anomaly detection, the system becomes inherently more robust against future, unknown spoofing methods.
Conclusion
RD-PAD shifts the paradigm of iris spoofing detection from binary classification to one-class anomaly detection. By leveraging the power of relativistic GANs, we create a detector that is not only competitive on known attacks but significantly more robust against the "unseen" threats that challenge modern biometric systems. Future work will investigate the application of this relativistic approach to other biometric modalities like face and fingerprint recognition.
Summary updated