Rank Correlation Measure: A Representational Transformation for Biometric Template Protection

Zhe Jin, Yen-Lung Lai, Andrew Beng Jin TeoharXiv 2016 · OpenReview

TL;DR

This paper introduces a novel biometric template protection method that transforms real-valued fingerprint features into compact index codes using a rank correlation-based approach. This technique aims to simultaneously achieve non-invertibility, cancellability, non-linkability, and performance preservation, addressing a key challenge in secure biometric systems.

What Problem Does This Paper Address?

Biometric systems, while convenient, store sensitive personal data as templates. Protecting these templates is crucial to prevent privacy breaches and identity theft, especially since biometric traits are immutable. Existing methods often struggle to satisfy all essential security criteria simultaneously:

  • Non-invertibility: The original template should be impossible to reconstruct from the protected version.
  • Cancellability: If a template is compromised, a new one can be generated from the same biometric trait without affecting the system.
  • Non-linkability/Diversity: Different instances of a protected template from the same individual should not be linkable across different applications.
  • Performance Preservation: The accuracy of the biometric system should not degrade significantly after template protection.

This paper tackles the challenge of developing a practical solution that balances these often-conflicting requirements for fingerprint biometrics.

Methods

The core of the proposed method is a ranking-based representational transformation inspired by "Winner Takes All" (WTA) hashing. Instead of directly using the numerical values of feature vectors, it focuses on their relative order. The process involves:

  1. Random Permutations: The input feature vector is permuted multiple times using different random seeds.
  2. Hardamard Product-Code Generation: For each set of permuted templates, a Hardamard product (element-wise multiplication) is computed. This step introduces non-linearity and makes inversion difficult.
  3. Index Code Generation: From the Hardamard product, the index of the maximum value within a sliding window (of size k) is recorded. This index, represented using log2(k) bits, forms the hashed code.
  4. Template Creation: Multiple independent hash functions (each with its own permutations and polynomial kernel degree p) are used to generate a final template of length m.

This approach transforms real-valued features into discrete index codes, making them resilient to minor numerical noise and inherently non-invertible due to the loss of absolute magnitude information.

Datasets

The proposed method was evaluated on publicly available fingerprint datasets:

  • FVC2002 (DB1, DB2, DB3)
  • FVC2004 (DB1, DB2, DB3)

These datasets are standard benchmarks for fingerprint recognition and verification.

Key Insights and Ideas

  • Rank Correlation is Key: The paper highlights that focusing on the relative order of feature components (rank correlation) rather than their absolute values is a powerful way to achieve template protection. This is more robust to noise and inherently harder to invert.
  • Non-linear Transformation for Security: The Hardamard product and the selection of the maximum within a window create a highly non-linear transformation, significantly increasing the difficulty of reconstructing the original feature vector.
  • Performance-Security Trade-off: The study demonstrates that increasing the complexity of the transformation (e.g., higher polynomial kernel degree p) can improve security but may slightly degrade accuracy, illustrating the common trade-off in cancellable biometrics.
  • Practicality: The method is designed to be computationally efficient and relatively simple to implement, making it suitable for practical applications.

Key Results

  • Accuracy: The proposed method achieved reasonable accuracy, with Equal Error Rates (EERs) comparable to or slightly higher than unprotected systems on benchmark datasets. For instance, on FVC2002 DB1, the proposed method achieved an EER of 0.43% with optimal parameters, compared to 0.59% for the unprotected baseline.
  • Security:
    • Non-invertibility: Security analysis showed that reconstructing the original feature vector would require an infeasible number of attempts (e.g., over 23,000 attempts for a single feature component on FVC2002 DB1).
  • ARM Attack Resistance: The transformation to rank space makes it computationally hard for Record Multiplicity (ARM) attacks to reconstruct the original template.
  • Cancellability: Experiments showed that newly generated templates from the same biometric trait are distinct enough to prevent imposter matching, satisfying the cancellability requirement.
  • Comparison: The proposed method outperformed existing minutiae-based template protection schemes on the tested datasets.

Limitations and Future Work

  • Lost Token/Seed Scenario: The paper notes that while user-specific permutation seeds enhance cancellability, the "lost token/seed" scenario (where seeds are lost) needs further focus, as it impacts accuracy, security, and privacy. The experiments were conducted with the same seed for all subjects to evaluate this scenario.
  • Generalizability: While the paper focuses on fingerprints, it suggests that the method could be adapted for other modalities with real-valued representations, and even binary representations by adapting the "first '1'" position.
  • Future Work: The authors imply that further exploration into optimizing parameters (m, k, p) for different biometric modalities and scenarios would be beneficial. Investigating the impact of different types of noise and perturbations on the rank correlation measure could also be a future direction.

Summary updated

Abstract

Despite a variety of theoretical-sound techniques have been proposed for biometric template protection, there is rarely practical solution that guarantees non-invertibility, cancellability, non-linkability and performance simultaneously. In this paper, a ranking-based representational transformation is proposed for fingerprint templates. The proposed method transforms a real-valued feature vector into index code such that the pairwise-order measure in the resultant codes are closely correlated with rank similarity measure. Such a ranking based technique offers two major merits: 1) Resilient to noises/perturbations in numeric values; and 2) Highly nonlinear embedding based on partial order statistics. The former takes care of the accuracy performance mitigating numeric noises/perturbations while the latter offers strong non-invertible transformation via nonlinear feature embedding from Euclidean to Rank space that leads to toughness in inversion. The experimental results demonstrate reasonable accuracy performance on benchmark FVC2002 and FVC2004 fingerprint databases, thus confirm the proposition of the rank correlation. Moreover, the security and privacy analysis justify the strong capability against the existing major privacy attacks.
BibTeX
@misc{jin2016rank,
  author = {Zhe Jin and Yen-Lung Lai and Andrew Beng Jin Teoh},
  title = {{Rank Correlation Measure: A Representational Transformation for Biometric Template Protection}},
  year = {2016},
  eprint = {1607.06902},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1607.06902},
}
Figure