Iris Presentation Attack Detection by Attention-based and Deep Pixel-wise Binary Supervision Network

Meiling Fang, Naser Damer, Fadi Boutros, Florian Kirchbuchner, Arjan KuijperarXiv 2021 · arXiv

TL;DR

Researchers have developed a new method for iris liveness detection that moves beyond simple "real vs. fake" labels by using Pixel-wise Binary Supervision (PBS) and spatial attention. By forcing the neural network to learn fine-grained local features and focus on the most discriminative parts of the eye, their A-PBS model significantly improves generalization across different sensors and unseen attack types, achieving state-of-the-art results on challenging benchmarks.

What Problem Does This Paper Address?

Iris recognition systems are increasingly used in security, but they are vulnerable to Presentation Attacks (PAs)—attempts to fool the system using high-quality printouts, digital replays, or textured contact lenses.

Current deep learning solutions for Presentation Attack Detection (PAD) typically treat the problem as a simple binary classification: is this image a "bona fide" (real) eye or an "attack"? While effective in controlled settings, this approach has three major flaws:

  1. Loss of Detail: Global binary labels ignore subtle, local cues (like the specific texture of a printed dot pattern) that distinguish a fake iris from a real one.
  2. Overfitting: Deep networks often "memorize" specific datasets rather than learning general features of liveness, causing them to fail when faced with a new camera or a new type of fake lens.
  3. Complexity: Many high-performing systems are too bulky or computationally expensive for deployment on mobile devices.

Methods

The authors propose a novel architecture based on DenseNet121, enhanced by two key innovations:

  1. Pixel-wise Binary Supervision (PBS): Instead of just one final "yes/no" output, the network is trained to produce a 2D map where every pixel is supervised. This forces the model to capture fine-grained, patch-level cues across the entire image rather than relying on a single global feature.
  2. Attention-based PBS (A-PBS): To prevent the model from wasting effort on irrelevant areas (like the pupil or the corners of the image), the authors added Spatial Attention Blocks. These modules learn to "weigh" different regions, automatically emphasizing the iris texture while suppressing noise from the background.
Figure 1. An overview of (a) baseline DenseNet, (b) proposed PBS and (c) proposed A-PBS networks.

The system uses a multi-scale approach, fusing features from low, middle, and high levels of the network to ensure it catches both coarse artifacts and microscopic texture anomalies.

Datasets

The model was evaluated on several prominent iris PAD benchmarks to test its robustness:

  • LivDet-Iris 2017: A competition dataset featuring cross-attack and cross-database scenarios (Clarkson, Notre Dame, and IIITD-WVU subsets).
  • NDCLD-2013 & NDCLD-2015: Focused on textured contact lens detection across different sensors (LG4000 and AD100).
  • IIIT-D CLI: A database specifically for cosmetic (textured) contact lens detection.

Key Insights and Ideas

  • Local vs. Global: The "aha moment" here is that liveness isn't a single global property; it's written in the local pixels. By supervising the network at the pixel level, it becomes much harder for the model to overfit to global image statistics.
  • Where Matters: The attention mechanism confirms that the most discriminative information for PAD resides in the circular iris region. As shown in the Score-CAM visualizations, the A-PBS model learns to ignore the pupil and the surrounding skin, focusing its "gaze" precisely where the attack artifacts (like lens patterns) are most visible.
Figure 3. Score-CAM visualizations for bona fide and attack sam- ples in the IIITD-WVU test set. The darker the color of the region, the higher the attention on this area. The column from left to right refers to the raw samples, maps produced by DenseNet, PBS, and A-PBS model, respectively. The first row is the bona fide samples and the left rows are textured contact lens and printouts attacks.
  • Generalization is the Goal: The A-PBS model showed significantly better performance in "cross-database" tests—where the model is trained on one set of eyes and tested on a completely different one—proving that it has learned generalizable features of iris liveness.

Key Results

The A-PBS model consistently outperformed state-of-the-art (SoTA) methods:

  • IIITD-WVU (Cross-DB): Achieved a Half Total Error Rate (HTER) of 6.50%, a significant improvement over previous methods like MSA (11.13%) and DenseNet (10.41%).
  • Notre Dame (Cross-PA): The HTER dropped to 3.94%, nearly half the error rate of the standard DenseNet baseline (8.14%).
  • Separability: As seen in the score distribution plots, A-PBS creates a much clearer separation between real eyes and attacks, leading to fewer false positives and false negatives.
Figure 2. Score distribution of bona fide (green) and PAs (blue) on the IIITD-WVU test set. The histograms from left to right are produced by DenseNet, PBS, and A-PBS, respectively. The larger separability (measured by Fisher Discriminant Ratio (FDR)) and maller overlap indicate higher classification performance.

Detailed comparisons against other algorithms can be found in the main results table: Comparison of A-PBS against SoTA algorithms on LivDet-Iris 2017.

Table 2. Iris PAD performance of our proposed methods and existing SoTA algorithms on LivDet-Iris 2017 databases in terms of APCER (%), BPCER (%) and HTER (%) which determined by a threshold of 0.5. The Winner in first column refers to the winner of each competition database. Bold numbers indicate the first two lowest HTERs.

Limitations and Future Work

While A-PBS shows excellent progress, the authors note that cross-database scenarios remain a major challenge for the entire field, with some error rates still hovering above 20% in the most difficult settings.

The near-perfect results on older datasets like NDCLD-2013 suggest those benchmarks may be reaching obsolescence. The authors call for the collection of larger-scale, more diverse iris PAD databases to continue pushing the boundaries of deep learning models and to better simulate real-world, "in-the-wild" conditions.

Summary updated

Abstract

Iris presentation attack detection (PAD) plays a vital role in iris recognition systems. Most existing CNN-based iris PAD solutions 1) perform only binary label supervision during the training of CNNs, serving global information learning but weakening the capture of local discriminative features, 2) prefer the stacked deeper convolutions or expert-designed networks, raising the risk of overfitting, 3) fuse multiple PAD systems or various types of features, increasing difficulty for deployment on mobile devices. Hence, we propose a novel attention-based deep pixel-wise binary supervision (A-PBS) method. Pixel-wise supervision is first able to capture the fine-grained pixel/patch-level cues. Then, the attention mechanism guides the network to automatically find regions that most contribute to an accurate PAD decision. Extensive experiments are performed on LivDet-Iris 2017 and three other publicly available databases to show the effectiveness and robustness of proposed A-PBS methods. For instance, the A-PBS model achieves an HTER of 6.50% on the IIITD-WVU database outperforming state-of-the-art methods.
BibTeX
@misc{fang2021iris,
  author = {Meiling Fang and Naser Damer and Fadi Boutros and Florian Kirchbuchner and Arjan Kuijper},
  title = {{Iris Presentation Attack Detection by Attention-based and Deep Pixel-wise Binary Supervision Network}},
  year = {2021},
  eprint = {2106.14845},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.14845},
}

Loading related content...

Figure