CIT-GAN: Cyclic Image Translation Generative Adversarial Network With Application in Iris Presentation Attack Detection

Shivangi Yadav, Arun RossWACV 2021

TL;DR

This paper introduces CIT-GAN, a novel Generative Adversarial Network architecture for synthesizing realistic iris presentation attack (PA) images across multiple domains. By employing a "Styling Network," CIT-GAN effectively learns domain-specific style characteristics, enabling the generation of diverse and high-quality synthetic PA samples that significantly improve the performance of iris presentation attack detection (PAD) systems, especially when training data is imbalanced.

What Problem Does This Paper Address?

Iris recognition systems are vulnerable to presentation attacks (PAs), where adversaries use fake irises (e.g., printed photos, artificial eyes, cosmetic contact lenses) to deceive the system. While many iris PAD methods exist, their performance is often hampered by a lack of sufficient and diverse training data, particularly for under-represented PA classes. Acquiring such comprehensive datasets is challenging and expensive. This paper addresses the critical need for generating realistic synthetic PA samples that can augment existing datasets and improve the robustness of iris PAD systems.

Methods

The core of this work is the Cyclic Image Translation Generative Adversarial Network (CIT-GAN). Unlike standard GANs, CIT-GAN is designed for multi-domain image translation. Its key components include:

  • Generator (G): Takes an input image and a domain label, translating it into an image with the style characteristics of the target domain.
  • Styling Network (S): This novel component learns domain-specific "style codes" from input images. These codes capture the unique visual attributes of each domain (e.g., the texture of a printed eye vs. a cosmetic lens). The Styling Network is trained to produce style codes that align with the correct domain, ensuring variability within each domain.
  • Discriminator (D): A multi-branch discriminator, where each branch is responsible for determining if an image is real or synthetic for a specific domain.

The translation process is driven by the style codes from the Styling Network, allowing the generator to synthesize images that mimic the appearance of specific PA types. The overall objective function incorporates adversarial loss, style loss, domain classification loss, and a cycle consistency loss to ensure that translated images retain essential content while adopting the target style.

Datasets

The authors utilized a combination of five existing iris PA datasets for training and testing:

  • Casia-iris-fake
  • Berc-iris-fake
  • NDCLD15
  • LivDet2017
  • MSU-IrisPA-01

These datasets contain bonafide iris images and various PA types, including cosmetic contact lenses, printed eyes, and artificial eyes. The images were pre-processed and cropped to a 256x256 resolution. The paper does not introduce a new benchmark dataset but focuses on generating synthetic data to augment existing ones.

Key Insights and Ideas

  • Domain-Specific Style Learning: The "Styling Network" is a crucial innovation. It allows CIT-GAN to learn and disentangle style characteristics specific to each PA domain, leading to more realistic synthetic samples than methods that treat all domains uniformly.
  • Synthetic Data for Robustness: The paper strongly demonstrates that synthetically generated PA samples, when used to balance imbalanced training datasets, can significantly improve the performance of existing iris PAD methods. This is a practical solution to the data scarcity problem.
  • Equivalence of Real and Synthetic PAs: Experiments show that PAD models trained with a mix of real and synthetic PAs, or even entirely with synthetic PAs, perform comparably to models trained solely on real PAs, validating the realism of CIT-GAN's output.
  • Multi-Domain Translation Efficiency: CIT-GAN's architecture is designed to handle multiple domains efficiently, overcoming limitations of previous methods that struggled with more than two domains.

Key Results

  • Image Quality: CIT-GAN achieved a superior Frechet Inception Distance (FID) score (average 32.79) compared to other multi-domain GANs like StarGAN (average FID 44.76) and StarGAN v2 (average FID 38.81), indicating higher quality synthetic images.
  • PAD Performance Improvement: In Experiment-4, where PAD methods were trained with balanced datasets including synthetic PAs, the True Detection Rate (TDR) at a 1% False Detection Rate (FDR) improved significantly. For instance, D-NetPAD's TDR increased from 92.54% to 97.89%, and BSIF+SVM's TDR improved from 28.11% to 51.11%.
  • Equivalence Validation: Experiments 2 and 3 showed that replacing real PA samples with synthetic ones resulted in only slight performance degradations for PAD methods, confirming the realism and utility of the generated data.

Limitations and Future Work

  • Limitations: While CIT-GAN shows excellent results, the paper notes that previous methods like StarGAN v2, though achieving higher FID scores than CIT-GAN in some cases, failed to fully capture the characteristics of certain PAs like artificial eyes. The paper implies that while CIT-GAN is superior, there's always room for improvement in capturing the nuances of all PA types.
  • Future Work: The authors propose extending their work to generate PA samples with more diverse variations and multi-domain style characteristics (e.g., "printed" cosmetic contact lenses). This would further enhance the ability of iris PAD methods to generalize across a wider range of presentation attacks, leading to more robust PA detection.

Summary updated

BibTeX
@inproceedings{yadav2021cit,
  author = {Shivangi Yadav and Arun Ross},
  title = {{CIT-GAN: Cyclic Image Translation Generative Adversarial Network With Application in Iris Presentation Attack Detection}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2021},
}
Figure