VISER: Visually-Informed System for Enhanced Robustness in Open-Set Iris Presentation Attack Detection

1. Problem Statement

Iris recognition is a high-security biometric modality vulnerable to Presentation Attacks (PAs), such as synthetic images, printed irises, or artificial eyes. While deep learning has improved Presentation Attack Detection (PAD), a critical challenge remains: Open-Set Robustness. Models often fail to generalize to unseen attack types (e.g., a model trained on printed attacks may fail on synthetic GAN-generated attacks).

To improve generalization, researchers utilize human saliency (maps indicating where humans look to make a decision) to guide model training. However, it is unclear which form of human saliency is optimal:

• Hand Annotations: Discrete, task-driven (e.g., mouse clicks/bounding boxes), but potentially noisy due to "motor-planning bottlenecks" and limited exploration.
• Eye Tracking: Continuous, physiological gaze data that captures reflexive attention but is susceptible to noise (micro-saccades, jitter).
• Foundation Models: Pre-trained models (e.g., DINOv2) that offer strong feature extraction but may lack specific domain adaptation for open-set iris PAD.

The paper aims to determine which saliency modality yields the best open-set performance and whether denoising eye-tracking data improves results.

2. Methodology

The authors propose VISER, a framework comparing various saliency-guided training approaches against a baseline and foundation models.

Experimental Setup

• Baseline: A DenseNet-121 architecture (D-NetPAD) trained with standard Cross-Entropy loss (XENT).
• Saliency-Guided Training: Models are trained using a combined loss function: Cross-Entropy + Mean Squared Error (MSE) between the model's Class Activation Map (CAM) and the target saliency map.
• Evaluation Paradigm: Leave-One-Attack-Type-Out. The model is trained on 6 attack types and tested on the 7th, repeated 12 times per attack type to ensure robust open-set evaluation.
• Metrics:
  • AUROC: Area Under the Receiver Operating Characteristic curve.
  • APCER @ 1% BPCER: Attack Presentation Classification Error Rate at a 1% Bonafide Presentation Classification Error Rate threshold (ISO/IEC 30107 standard).

Saliency Modalities Tested

1. Segmentation Masks: Generated via open-source iris recognition tools (least intricate).
2. Hand Annotations: Derived from a study by Boyd et al. [5]. To match the entropy of eye-tracking data, three variations were created using Gaussian blur:
   • Low Entropy: Original annotations.
   • Equal Entropy: Blur kernel size 5.
   • High Entropy: Blur kernel size 10.
3. Eye Tracking Heatmaps: Derived from Dowling et al. [17], capturing gaze from experts and non-experts.
   • Full Eye Tracking: Gaze over the entire evaluation window.
   • Initial Eye Tracking: Gaze during the participant's first impression.
   • De-noised Variants: Raw eye-tracking data processed with HDBSCAN (minimum cluster size 5) to remove outlier fixations (noise) before aggregation.
4. Foundation Models: DINOv2-Base embeddings extracted and fed into three classifiers: Logistic Regression (LogReg), Linear SVM, and RBF-SVM.

3. Key Contributions

1. Comprehensive Saliency Comparison: The first study to directly compare hand annotations, eye-tracking heatmaps, segmentation masks, and foundation model embeddings for open-set iris PAD.
2. Denoising Pipeline: Introduction of an HDBSCAN-based denoising method for raw eye-tracking data, demonstrating its efficacy in improving saliency-guided training.
3. VISER Framework: A validated system showing that denoised initial eye-tracking heatmaps provide the most significant generalization improvements over traditional cross-entropy baselines.
4. Open-Source Release: Provision of trained models, code, and saliency maps to facilitate reproducibility.

4. Key Results

The results are presented as improvements (or deltas) over the DenseNet XENT baseline.

• RQ1: Eye Tracking vs. Other Saliency:

  • Eye tracking significantly outperformed hand annotations and segmentation masks.
  • De-noised Initial Eye Tracking achieved the highest improvement: +0.0608 AUROC and -0.1063 APCER (lower is better).
  • Hand annotations and segmentation masks generally failed to surpass the baseline in APCER metrics.

• RQ2: Saliency-Guided vs. Foundation Models:

  • Saliency-guided training (specifically with eye tracking) outperformed foundation model approaches (DINOv2 + Classifiers) in open-set scenarios.
  • Among foundation models, only DINOv2 + SVM-RBF showed marginal AUROC improvement (+0.0485), while LogReg and Linear SVM failed to improve the baseline.
  • Foundation models showed inconsistent performance across different attack types, whereas eye-tracking methods showed broad improvements.

• RQ3: Denoising Impact:

  • Denoising improved Initial Eye Tracking performance significantly.
  • However, denoising Full Eye Tracking maps actually resulted in lower performance compared to the raw Full Eye Tracking maps, suggesting that the "noise" in the full dataset contained useful information or that the clustering parameters were less optimal for the full duration.

• Specific Attack Type Observations:

  • Eye tracking methods showed massive improvements on "Artificial" (prosthetic eyes) and "Printout" attacks.
  • The "Contacts + Print" attack type was already well-handled by the baseline, leaving little room for improvement.

5. Significance and Conclusion

• Superiority of Physiological Signals: The study confirms that eye-tracking saliency (a first-order physiological signal) is superior to hand annotations (a second-order motor signal) for guiding deep learning in iris PAD. This supports the hypothesis that hand annotations suffer from a "motor-planning bottleneck," where users may not highlight all subtle visual cues they perceive.
• Optimal Training Strategy: The best approach for open-set iris PAD is to use denoised initial gaze heatmaps to guide the training of a DenseNet architecture. This approach yields state-of-the-art robustness against unseen attack types.
• Limitations of Foundation Models: While foundation models like DINOv2 are powerful, their performance in open-set iris PAD is inconsistent and often inferior to carefully curated saliency-guided training, likely due to unknown training data overlaps and lack of specific domain adaptation.
• Future Work: The authors suggest that while denoising is beneficial for initial gaze, future research should explore different clustering parameters for full-duration eye tracking and validate these findings in other biometric domains.

In summary, VISER demonstrates that leveraging cleaned, initial human gaze data provides a more robust and generalizable signal for training iris PAD systems than traditional annotations or off-the-shelf foundation models.