DeFraudNet:End2End Fingerprint Spoof Detection using Patch Level Attention

Anusha B, Sayan Banerjee, Subhasis ChaudhuriWACV 2020

TL;DR

This paper introduces DeFraudNet, a novel end-to-end deep learning architecture for fingerprint spoof detection that leverages both global image features and local patch-level information. By employing a unique patch attention mechanism, DeFraudNet effectively identifies discriminative patches and fuses their information with global context, achieving state-of-the-art performance across various datasets and challenging cross-sensor/cross-material scenarios.

What Problem Does This Paper Address?

Fingerprint recognition systems are widely used for security, but they are vulnerable to spoofing attacks where fake fingerprints are presented to fool the sensor. While significant progress has been made, detecting sophisticated spoofs, especially those made from diverse materials and across different sensor types, remains a challenge. Existing methods often struggle with cross-material and cross-sensor generalization, limiting their real-world applicability. This paper aims to address this gap by developing a more robust and generalized spoof detection system.

Methods

The DeFraudNet architecture is an end-to-end system that combines global and local feature extraction. The core components include:

  • Preprocessing: Input grayscale fingerprint images are converted into 3-channel images by incorporating Local Binary Patterns (LBP) and Gabor filter responses. This enriches the input with textural and edge information, which is crucial for distinguishing live from fake fingerprints.
  • DenseNet Backbone: Two DenseNets are employed. A larger DenseNet (40 layers) processes the entire image for global features, while a smaller DenseNet (10 layers) is used to extract features from individual patches. DenseNets are chosen for their efficiency and ability to propagate features effectively.
  • Attention Mechanisms:
    • Channel and Spatial Attention: These modules are applied to the patch features to highlight the most relevant information within each patch.
  • Patch Attention Network: This novel component is the key innovation. It learns to identify the most discriminative patches from a set of extracted patches and assigns them importance scores. This allows the network to focus on the most informative parts of the fingerprint for classification.
  • Feature Fusion and Classification: The attention-weighted patch features are fused with the global image features. This fused representation is then passed through a softmax layer for final live/fake classification. The patch attention mechanism provides an efficient fusion strategy, learning to aggregate information based on patch importance, making it robust to noise and errors.

Datasets

The proposed method is evaluated on four publicly available datasets from the Liveness Detection Competition (LivDet):

  • LivDet 2011
  • LivDet 2013
  • LivDet 2015
  • LivDet 2017

These datasets contain fingerprint images acquired from various sensors and using different spoof materials, allowing for comprehensive evaluation of cross-sensor, cross-material, and cross-dataset performance.

Key Insights and Ideas

  • Patch-level attention is crucial for generalization: The paper highlights that focusing on specific, discriminative patches, rather than just global features or all patches equally, significantly improves performance, especially in challenging cross-scenario tests.
  • Combining global and local features is powerful: The synergy between features extracted from the entire image and those from selected local patches leads to a more robust representation that can better handle variations in spoof materials and sensor types.
  • End-to-end learning simplifies the pipeline: By learning patch selection and feature fusion within a single network, the method avoids complex multi-stage preprocessing steps often required by previous approaches, making it more efficient and easier to deploy.
  • DenseNet's efficiency is a practical advantage: The choice of DenseNet as the backbone contributes to reduced memory usage and computational cost, making the system potentially suitable for deployment on resource-constrained devices like IoT devices.

Key Results

The DeFraudNet achieves impressive results, outperforming the current state-of-the-art on several benchmarks:

  • LivDet 2017: Average accuracy of 99.52%.
  • LivDet 2015: Average accuracy of 99.16%.
  • LivDet 2011: Average accuracy of 99.72%.
  • State-of-the-art improvement: Outperforms previous methods by 3% and 4% on LivDet 2015 and 2011, respectively.
  • Cross-sensor and cross-material performance: The method demonstrates significant improvements in these challenging scenarios, with a striking 85% decrease in average classification error for some cross-material tests compared to previous state-of-the-art approaches.

Limitations and Future Work

  • Cross-sensor performance on LivDet 2013: While generally strong, the paper notes that cross-sensor metrics for LivDet 2013 were an exception where the proposed network did not outperform the state-of-the-art.
  • Scope for improvement: The authors acknowledge that there is still considerable scope for improvement in cross-sensor and cross-dataset classification errors.
  • Future integration: The authors plan to integrate DeFraudNet with fingerprint authentication and recognition systems to create a complete system with built-in spoof detection.

Summary updated

BibTeX
@inproceedings{b2020defraudnet,
  author = {Anusha B and Sayan Banerjee and Subhasis Chaudhuri},
  title = {{DeFraudNet:End2End Fingerprint Spoof Detection using Patch Level Attention}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2020},
}
Figure