Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Privacy-Safe Iris Presentation Attack Detection

Mahsa Mitcheff, Patrick Tinsley, Adam Czajka

TL;DR

The paper tackles privacy and data collection challenges in iris presentation attack detection by proposing a four-step framework that trains PAD models exclusively on synthetically-generated, identity-leakage-free iris images. It uses two StyleGAN2-ADA based generators to synthetically produce iris images with and without textured contact lenses and includes a post-hoc identity leakage mitigation step to ensure no subject-identifying information is leaked. The results show that synthetic-data-trained PAD models achieve an average AUROC of about $0.90$ to $0.93$, compared with roughly $0.97$ when trained on authentic data, indicating a measurable but modest gap that may narrow with higher fidelity synthesis. The work demonstrates the feasibility of privacy-safe iris PAD and provides code and models to support further research and potential practical deployment.

Abstract

This paper proposes a framework for a privacy-safe iris presentation attack detection (PAD) method, designed solely with synthetically-generated, identity-leakage-free iris images. Once trained, the method is evaluated in a classical way using state-of-the-art iris PAD benchmarks. We designed two generative models for the synthesis of ISO/IEC 19794-6-compliant iris images. The first model synthesizes bona fide-looking samples. To avoid ``identity leakage,'' the generated samples that accidentally matched those used in the model's training were excluded. The second model synthesizes images of irises with textured contact lenses and is conditioned by a given contact lens brand to have better control over textured contact lens appearance when forming the training set. Our experiments demonstrate that models trained solely on synthetic data achieve a lower but still reasonable performance when compared to solutions trained with iris images collected from human subjects. This is the first-of-its-kind attempt to use solely synthetic data to train a fully-functional iris PAD solution, and despite the performance gap between regular and the proposed methods, this study demonstrates that with the increasing fidelity of generative models, creating such privacy-safe iris PAD methods may be possible. The source codes and generative models trained for this work are offered along with the paper.

Privacy-Safe Iris Presentation Attack Detection

TL;DR

The paper tackles privacy and data collection challenges in iris presentation attack detection by proposing a four-step framework that trains PAD models exclusively on synthetically-generated, identity-leakage-free iris images. It uses two StyleGAN2-ADA based generators to synthetically produce iris images with and without textured contact lenses and includes a post-hoc identity leakage mitigation step to ensure no subject-identifying information is leaked. The results show that synthetic-data-trained PAD models achieve an average AUROC of about to , compared with roughly when trained on authentic data, indicating a measurable but modest gap that may narrow with higher fidelity synthesis. The work demonstrates the feasibility of privacy-safe iris PAD and provides code and models to support further research and potential practical deployment.

Abstract

This paper proposes a framework for a privacy-safe iris presentation attack detection (PAD) method, designed solely with synthetically-generated, identity-leakage-free iris images. Once trained, the method is evaluated in a classical way using state-of-the-art iris PAD benchmarks. We designed two generative models for the synthesis of ISO/IEC 19794-6-compliant iris images. The first model synthesizes bona fide-looking samples. To avoid ``identity leakage,'' the generated samples that accidentally matched those used in the model's training were excluded. The second model synthesizes images of irises with textured contact lenses and is conditioned by a given contact lens brand to have better control over textured contact lens appearance when forming the training set. Our experiments demonstrate that models trained solely on synthetic data achieve a lower but still reasonable performance when compared to solutions trained with iris images collected from human subjects. This is the first-of-its-kind attempt to use solely synthetic data to train a fully-functional iris PAD solution, and despite the performance gap between regular and the proposed methods, this study demonstrates that with the increasing fidelity of generative models, creating such privacy-safe iris PAD methods may be possible. The source codes and generative models trained for this work are offered along with the paper.
Paper Structure (26 sections, 1 equation, 6 figures, 2 tables)

This paper contains 26 sections, 1 equation, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Iris Image Synthesis
  4. Iris Presentation Attack Detection
  5. Synthetic Data in the Context of PAD
  6. Methodology
  7. Generative Model Selection
  8. Overview of the Proposed Framework
  9. Training Generative Models on Authentic Data (Step 1 in Fig. \ref{['fig:pipeline']})
  10. Datasets Used
  11. Authentic TCL:
  12. Authentic noTCL:
  13. Training
  14. Image Synthesis and Identity Leakage Mitigation (Step 2 in Fig. \ref{['fig:pipeline']})
  15. Synthesis
  16. ...and 11 more sections

Figures (6)

  • Figure 1: Approaches to source training samples to develop iris presentation attack detection methods: (a) a classical approach, in which authentic images of living individuals and physical attacks are used, (b) similar to (a) but targeting specifically detection of synthetically-generated iris images, or "deep fakes", (c) a newer approach, in which bona fide samples are synthesized, and then used to carry out physical presentation attacks captured by iris sensors, and (d) the approach proposed in this paper, in which both bona fide and attack samples are synthesized and used to train iris presentation attack detection models. BF and PA stand for "bona fide" and "presentation attack," respectively. A person icon next to the dataset icon denotes the presence of identity information in the data.
  • Figure 2: The pipeline of privacy-safe, synthetic data-only iris presentation attack detection (PAD) training and validation. TCL and noTCL denote images of irises with and without contact lenses, respectively. After training generative models ( Step 1), we exclusively use synthetically-generated data (mimicking irises both with and without textured contact lenses) to train iris PAD as usual ( Step 3). The iris matcher is used (in Step 2) to exclude synthetic samples that are "too close" to non-synthetic samples used for generative models training, which prevents the "leakage" of identity information from the training set into the generated samples. Resulting iris PAD methods are tested on regular (non-synthetic) data composed of bona fide and fake samples ( Step 4).
  • Figure 3: Examples of authentic (upper row) and synthetically-generated by a conditional StyleGAN2-ADA (bottom row) samples without textured contact lens (noTCL) and with textured contact lens (TCL) of a given brand.
  • Figure 4: Distributions of the ISO/IEC 29794-6 overall quality score for iris images without contact lenses (noTCL) shown on the left plot, and for iris images with contact lenses (TCL) shown on the right plot. The quality score ranges from 0 to 100, with higher scores indicating better image quality.
  • Figure 5: Average DET curves (thick lines), with shaded areas representing one standard deviation from five train-test runs, obtained for both experiments (E1 in blue and E2 in red) and for all three model backbones.
  • ...and 1 more figures