A Novel Anti-Spoofing Solution for Iris Recognition Toward Cosmetic Contact Lens Attack Using Spectral ICA Analysis

Sheng-Hsun Hsieh, Yung-Hui Li, Wei Wang, Chung-Hao TienSensors 2018 · OpenReview

TL;DR

This paper introduces a novel anti-spoofing method for iris recognition that specifically targets cosmetic contact lens (CCL) attacks. By employing a dual-band spectral imaging system and Independent Component Analysis (ICA), the approach effectively separates the natural iris texture from the CCL pattern, enabling accurate recognition even when CCLs are worn. This significantly improves the False Rejection Rate (FRR) from 10.52% to 0.57% in their experimental setup.

What Problem Does This Paper Address?

Iris recognition is a highly accurate biometric modality, but like all biometrics, it's vulnerable to spoofing. A common and challenging spoofing method involves wearing cosmetic contact lenses (CCLs). These lenses, often designed to mimic natural iris patterns, can obscure or alter the genuine iris texture, leading to degraded recognition performance or outright failures. Existing anti-spoofing techniques often focus on liveness detection (is the eye real?) or explicit CCL detection, which may require user cooperation or fail to recognize novel CCL designs. This paper tackles the problem of performing iris recognition directly on a subject wearing a CCL, without needing to remove it, by leveraging the spectral properties of the iris and CCLs.

Methods

The core of the proposed solution lies in a combination of hardware and software:

  • Dual-Band Spectral Imaging System: The system uses a camera equipped with two sensors and distinct spectral filters (one low-pass and one high-pass) to capture two images of the same iris under illumination. These two images represent mixed signals where the natural iris texture and the CCL pattern are superimposed, but with different spectral mixing ratios.
  • Independent Component Analysis (ICA): ICA is a blind source separation technique used to disentangle statistically independent signals from a mixture. The authors assume that the natural iris texture and the CCL pattern are statistically independent in the spectral domain. By applying ICA to the two spectral images, they aim to separate them into their individual components: one representing the natural iris and the other the CCL.
  • Mask Generation and Feature Extraction: After ICA separates the components, a binary mask is generated to isolate the natural iris texture. This is achieved by analyzing the characteristics of the estimated components (e.g., histogram distribution, negentropy) to distinguish between the iris and the CCL. The CCL component is identified, and a mask is created to remove its influence from the mixed image, effectively restoring a "CCL-free" iris image for subsequent recognition.
  • Iris Recognition Algorithm: The paper utilizes a standard iris recognition algorithm, inspired by Daugman's work, which involves iris segmentation, normalization, feature extraction (using Log-Gabor filters to create an iris code), and matching via Hamming Distance (HD).

Datasets

The study used a custom dataset for proof-of-concept testing:

  • Dataset: 10 East Asian individuals, each with two irises, resulting in 20 iris classes.
  • Data Capture: For each iris class, 10 enrollment images were captured (naked eye). Additionally, 10 image pairs were captured in both spectral channels (visible and NIR) for testing. This resulted in a total of 200 enrollment images and 200 test image pairs.
  • CCL Type: Subjects wore a specific type of contact lens (TICON) with a circular band pattern formed by random dots.

Key Insights and Ideas

  • Spectral Distinctness as a Key: The fundamental idea is that natural irises and artificial CCLs, despite their visual similarity, exhibit different spectral reflectance properties. This spectral difference, combined with statistical independence, is the basis for separating them.
  • ICA for Source Separation: ICA is a powerful tool for disentangling mixed signals when the underlying sources are statistically independent. The paper demonstrates its effectiveness in separating the spectral components of iris and CCL.
  • Recognition-Centric Approach: Unlike many anti-spoofing methods that focus on detecting a CCL, this approach aims to recognize the iris even in the presence of a CCL. This is a significant shift in focus.
  • Hardware-Software Synergy: The solution elegantly combines specialized hardware (dual-band camera) with advanced software (ICA) to achieve its goal.
  • Ambiguity Handling: The paper acknowledges and addresses inherent ambiguities in ICA (scaling and order) by leveraging image properties like histograms and negentropy to correctly identify the iris and CCL components.

Key Results

  • Baseline Performance (Naked Eye): The iris recognition system achieved a high Sensitivity Index (SI) of 8.82, with a False Rejection Rate (FRR) of 0% at a False Acceptance Rate (FAR) of 0.1%, and an Equal Error Rate (EER) of 0%.
  • Performance with CCLs (Without ICA): When subjects wore CCLs, the FRR dramatically increased to 10.52%, and the EER rose to 1.94%. This highlights the significant impact of CCLs on standard iris recognition.
  • Performance with CCLs (With ICA): After applying the proposed ICA-based anti-spoofing scheme, the FRR was reduced to a mere 0.57%, and the EER to 0.26%. This represents an 18.5x reduction in FRR compared to the baseline without ICA.
  • Qualitative Results: The generated masks effectively isolated the CCL pattern (red in figures) and the natural iris texture (blue in figures), demonstrating the successful separation of components.

Limitations and Future Work

  • Limited CCL Types: The study focused on a specific type of CCL with a random dot pattern. The authors acknowledge that the effectiveness might vary with different CCL designs and colors.
  • Number of Channels: The current system uses two spectral channels. For multi-colored CCLs, more channels or a more complex ICA approach might be necessary to disentangle multiple spectral primitives.
  • Spectral Distinctness: The paper notes that the effectiveness of the ICA-based scheme is expected to increase when the spectral distinctness between CCL and natural iris is greater.
  • Future Work: The authors suggest investigating scenarios with multiple CCL colors, exploring spectral analysis differences across different iris races and CCL pigments, and further optimizing the setup for broader applicability. They also mention the need to balance countermeasure capability with computational complexity for real-time applications.

Summary updated

Abstract

In this study, we maneuvered a dual-band spectral imaging system to capture an iridal image from a cosmetic-contact-lens-wearing subject. By using the independent component analysis to separate individual spectral primitives, we successfully distinguished the natural iris texture from the cosmetic contact lens (CCL) pattern, and restored the genuine iris patterns from the CCL-polluted image. Based on a database containing 200 test image pairs from 20 CCL-wearing subjects as the proof of concept, the recognition accuracy (False Rejection Rate: FRR) was improved from FRR = 10.52% to FRR = 0.57% with the proposed ICA anti-spoofing scheme.
BibTeX
@inproceedings{hsieh2018novel,
  author = {Sheng-Hsun Hsieh and Yung-Hui Li and Wei Wang and Chung-Hao Tien},
  title = {{A Novel Anti-Spoofing Solution for Iris Recognition Toward Cosmetic Contact Lens Attack Using Spectral ICA Analysis}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=O8Hre5Eqozu},
}
Figure