Vikriti-ID: A Novel Approach for Real Looking Fingerprint Data-Set Generation

Rishabh Shukla, Aditya Sinha, Vansh Singh, Harkeerat KaurWACV 2024

TL;DR

This paper introduces Vikriti-ID, a novel generative model that creates realistic synthetic fingerprint datasets with multiple impressions per identity. This addresses the critical shortage of publicly available, large-scale fingerprint data, enabling better training and evaluation of fingerprint recognition systems.

What Problem Does This Paper Address?

Automatic fingerprint recognition systems are widely used in various applications, from mobile security to border control. However, their development and evaluation are significantly hampered by the scarcity of large, publicly accessible fingerprint databases. Existing datasets often lack sufficient identities and multiple impressions per identity, which is crucial for training robust algorithms and understanding intra-user variability. This limitation hinders progress in areas like large-scale identity management systems. Furthermore, privacy concerns and regulations make it difficult to create and share new real-world fingerprint datasets.

Methods

Vikriti-ID employs a multi-stage generative approach combining Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). The process begins with a VAE that takes a random noise matrix and generates an intermediate image, capturing semantic fingerprint characteristics. This intermediate image is then fed into a GAN-based generator, which uses class-specific keys to produce a unique fingerprint identity. Finally, an impression generator module creates multiple, distinct impressions for each generated identity, ensuring intra-user variation while maintaining identity consistency. This approach is designed to overcome limitations of previous methods, such as unrealism and limited variation, by incorporating domain knowledge and generating class-specific samples.

Datasets

The authors generated a large synthetic dataset named Vikriti-ID, comprising 500,000 unique fingerprints, each with 10 associated impressions, across five classes (Arch, Left Loop, Right Loop, Tented Arch, Whorl). This dataset was used for training and evaluation. For fine-tuning and comparison, they also utilized publicly available datasets: FVC (Fingerprint Verification Competition) datasets (DB1, DB2, DB3, DB4 from FVC2000 and FVC2002) and SOCOFing.

Key Insights and Ideas

  • Addressing the data bottleneck: The core insight is that synthetic data generation, when done effectively, can significantly alleviate the data scarcity problem in biometrics, enabling research that was previously infeasible.
  • Multi-stage generation for realism: Breaking down the generation process into VAE for intermediate representation and GAN for identity synthesis, followed by an impression generator, allows for better control over realism and variation.
  • Class-specific keys for identity control: Using class-specific keys in the GAN stage helps ensure that generated fingerprints adhere to distinct class characteristics, contributing to both realism and uniqueness.
  • Generating multiple impressions is key: The ability to generate multiple impressions for a single identity is crucial for simulating real-world scenarios and training models that can handle intra-user variations.
  • No data leakage: A critical finding is that the generated dataset does not appear to "leak" identity information from its training data, which is essential for privacy and secure sharing.

Key Results

The Vikriti-ID generated dataset, when used to train a deep network (inspired by [13]), achieved an impressive Equal Error Rate (EER) of 0.16% and an Area Under the Curve (AUC) of 0.89%. This performance is comparable to or better than results obtained using publicly available datasets. The paper also demonstrates that a model trained on Vikriti-ID data and then fine-tuned on the FVC2002 dataset achieved an EER of 0.16% and an AUC of 0.9. Quantitative evaluations using MCC matcher showed good score separation between genuine and imposter matches, supporting the realism and uniqueness of the generated fingerprints.

Limitations and Future Work

The authors acknowledge that while their synthetic fingerprints are realistic, they may still exhibit a "domain gap" compared to real-world fingerprints. Future work suggested includes refining the generation techniques using advanced GAN architectures, exploring multi-modal fusion for enhanced accuracy, and conducting real-world testing to assess robustness against adversarial attacks and generalization to different domains. They also emphasize the importance of continuously updating the Vikriti-ID database and sharing it with the research community.

Summary updated

BibTeX
@inproceedings{shukla2024vikriti,
  author = {Rishabh Shukla and Aditya Sinha and Vansh Singh and Harkeerat Kaur},
  title = {{Vikriti-ID: A Novel Approach for Real Looking Fingerprint Data-Set Generation}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
}
Figure