PaperlayerView exact revision on arXiv

Source-linked AI summary

Fingerprint Spoof Buster

Tarang Chugh, Kai Cao, Anil K. Jain

arXiv:1712.04489v1cs.CV

TL;DR

Fingerprint spoof attacks threaten the security of authentication systems, motivating accurate and generalizable detection methods. The paper trains MobileNet-v1 CNNs on minutiae-centered, aligned local patches and reports reduced error rates across multiple evaluation settings, while also supporting local visualization of spoof regions.

  • Problem

    The paper addresses the need for accurate and generalizable fingerprint spoof detection because spoof attacks can compromise authentication-system security.

  • Method

    The approach extracts minutiae-centered and orientation-aligned local patches, trains MobileNet-v1 CNN models, and averages patch-level spoofness scores for testing.

  • Results

    The approach reduces error rates relative to the state of the art in intra-sensor, cross-material, cross-sensor, and cross-dataset evaluations.

  • Takeaways & Limitations

    The patch-based representation provides fine-grained spoof cues and highlights fingerprint regions as live or spoof for operator examination.

Abstract

from arXiv · show

The primary purpose of a fingerprint recognition system is to ensure a reliable and accurate user authentication, but the security of the recognition system itself can be jeopardized by spoof attacks. This study addresses the problem of developing accurate, generalizable, and efficient algorithms for detecting fingerprint spoof attacks. Specifically, we propose a deep convolutional neural network based approach utilizing local patches centered and aligned using fingerprint minutiae. Experimental results on three public-domain LivDet datasets (2011, 2013, and 2015) show that the proposed approach provides state-of-the-art accuracies in fingerprint spoof detection for intra-sensor, cross-material, cross-sensor, as well as cross-dataset testing scenarios. For example, in LivDet 2015, the proposed approach achieves 99.03% average accuracy over all sensors compared to 95.51% achieved by the LivDet 2015 competition winners. Additionally, two new fingerprint presentation attack datasets containing more than 20,000 images, using two different fingerprint readers, and over 12 different spoof fabrication materials are collected. We also present a graphical user interface, called Fingerprint Spoof Buster, that allows the operator to visually examine the local regions of the fingerprint highlighted as live or spoof, instead of relying on only a single score as output by the traditional approaches.

II. PROPOSED SPOOF DETECTION APPROACH

The proposed system trains MobileNet-v1 models on minutiae-based local patches and averages patch-level spoofness scores during testing. The study evaluates this approach using LivDet datasets and newly collected fingerprint presentation-attack datasets.

  • The approach uses offline minutiae detection, aligned local-patch extraction, MobileNet training, and online averaging of patch-level spoofness scores.The final spoof decision is based on a threshold learned from the training data.
  • The study evaluates LivDet 2011, 2013, and 2015 datasets, each containing more than 16,000 images from four fingerprint readers.The datasets contain equal numbers of live and spoof fingerprints split between training and testing sets.

2) MSU Fingerprint Presentation Attack Dataset:

The study collects the MSU-FPAD and PBSKD datasets using two fingerprint readers and diverse spoof fabrication materials. These datasets include live and spoof images intended to support fingerprint presentation-attack research.

  • MSU-FPAD contains 9,000 live images and 10,500 spoof images captured using CrossMatch Guardian 200 and Lumidigm Venus 302 readers.Its spoof materials include Ecoflex, PlayDoh, 2D prints on matte paper, and 2D prints on transparency film.
  • PBSKD includes 900 spoof fingerprint images fabricated from 10 different spoof materials.The listed materials include Ecoflex, gelatin, latex body paint, coatings, modeling material, wood glue, liquid latex, and 2D prints.
  • PBSKD contains 1,800 fingerprint images in total, with example images acquired using the CrossMatch Guardian 200 and Lumidigm Venus 302 readers.The Lumidigm reader does not image silicone spoofs with NanoTips and BarePaint coatings.

B. Minutiae Detection

Fingerprint images are standardized and processed around detected minutiae to produce aligned local patches. These patches are cropped at selected resolutions and used as CNN inputs.

  • Fingerprint images from the LivDet datasets are resized to 500 dpi before minutiae extraction to standardize resolution.The minutiae detector used was designed for 500 dpi images.
  • Each detected minutia is represented by spatial coordinates and orientation, which determine the center and alignment of its corresponding local patch.The local patches are extracted around all detected minutiae.
  • After rotation-based alignment, the central p × p region is cropped from each larger patch and used to train the CNN.The default cropped patch size is p = 96, while the larger patch prevents information loss during alignment.
  • Multi-resolution experiments use patch sizes p ∈ {64, 96, 128}, with all patches resized to 224 × 224 for MobileNet-v1.Independent CNN models are trained for the different patch sizes and can be fused.

D. MobileNet CNN

The method uses MobileNet-v1 to classify aligned fingerprint patches as live or spoof. The architecture is selected for lower computational cost and latency while supporting patch-level visualization and global spoofness scoring.

  • MobileNet-v1 uses depth-wise separable convolutions and has fewer parameters and lower latency than the compared VGG and Inception-v3 architectures.The passage reports 4.24M parameters and 100ms classification time for MobileNet-v1, compared with 23.2M and 800ms for Inception-v3.
  • The ImageNet 1000-unit softmax layer is replaced with a 2-unit softmax layer for live-versus-spoof classification.The network is implemented using TF-Slim.
  • Training uses RMSProp with asynchronous gradient descent, batch size 100, and augmentation including brightness adjustment, random cropping, and vertical flipping.Separate networks are trained for multi-resolution patches using the same parameters.

E. Fine-grained Fingerprint Image Representation

The approach uses minutiae-based local patches to provide fine-grained fingerprint representations, especially for partial sensor regions and partial spoofs. Patch-level spoofness scores are aggregated into an image-level decision using a learned threshold, while visualizations expose the regions supporting live or spoof classifications.

  • Fine-grained representation: Patch-based detection provides a fine-grained representation suited to partial spoofs and small sensors capturing regions typically around 150 × 150.Whole-fingerprint detectors can be ineffective when spoofing conceals only a limited region of a live finger.
  • Patch-level scoring: Minutiae-based local patches are classified by a MobileNet-v1 model whose softmax output defines a Spoofness Score in [0−1].Higher scores indicate greater support that an input patch belongs to the spoof class.
  • Score aggregation: Patch-level spoofness scores are averaged across the k extracted patches, and multi-resolution scores for patch sizes 64, 96, and 128 are averaged into a final score.The resulting score aggregates local evidence across patches and resolutions.
  • Classification threshold: The learned threshold minimizes average training classification error and performs slightly better than a predefined threshold of 0.5.Scores below the threshold are classified as live, while scores at or above it are classified as spoof.
  • Visualization and robustness: A modified spoof patch retains a high spoofness score after artifact removal, indicating robustness of the proposed approach.The figure visualizes live, spoof, and modified spoof patch embeddings in a 1024-dimensional feature space and corresponding heat maps.

G. On Robustness of Patch-based Representation

The study examines whether local-patch representations remain robust when fingerprint spoof artifacts are altered or removed. Its evaluation uses LivDet performance metrics and reports spoof breaches at a fixed legitimate-user reject rate.

  • G. On Robustness of Patch-based Representation: The robustness analysis tests whether patch-based features capture discriminatory information from local regions such as missing ridges and air bubbles.The analysis uses MobileNet-v1 bottleneck representations of local patches to inspect this behavior.
  • G. On Robustness of Patch-based Representation: 0.00 and 0.99 are the reported spoofness scores for the live and spoof patches shown in the robustness analysis.The spoof patch was modified with GIMP to remove visible artifacts before evaluation.
  • G. On Robustness of Patch-based Representation: ACE is evaluated as the sum of the percentages of misclassified live and spoof fingerprints.Ferrlive measures misclassified live fingerprints, while Ferrfake measures misclassified spoof fingerprints.
  • G. On Robustness of Patch-based Representation: Ferrfake @ Ferrlive = 1% measures the percentage of spoofs breaching security when the legitimate-user reject rate is at most 1%.The study reports this metric for each experiment.

B. Results

The results evaluate the proposed approach across intra-sensor, cross-material, cross-sensor, and cross-dataset settings using LivDet and additional presentation-attack datasets. Experiments show benefits from minutiae-centered patches, 96 × 96 patches, and multi-resolution score fusion.

  • B. Results: The evaluation covers four scenarios reflecting robustness to new spoof materials, different sensors, and different data-collection environments.The reported datasets include LivDet 2011, 2013, and 2015, MSU-FPAD, and PBSKD.
  • B. Results: Intra-sensor experiments use the same sensor for training and testing, with test spoof materials known during training.The proposed MobileNet-v1 trained from scratch on minutiae-based patches outperformed fine-tuning, while fusion with a whole-image model added no performance improvement.
  • B. Results: ACE decreased from 1.24% to 0.88%, and Ferrfake decreased from 1.41% to 0.58% @ Ferrlive = 1% when three patch resolutions were fused.The comparison used Biometrika images from LivDet 2011.
  • B. Results: The proposed approach produced correctly and incorrectly classified live and spoof fingerprints across the evaluated sensor examples.The examples include the Biometrika sensor from LivDet 2015.
  • B. Results: Minutiae-centered patches achieved a 78% reduction in average classification error, compared with 33% for randomly sampled patches.The comparison used models trained on a similar number of local patches from LivDet 2015.
  • B. Results: Local features outperformed whole-image features, minutiae-selected patches outperformed random patches, 96 × 96 patches performed best individually, and multi-resolution fusion improved performance.These findings summarize the patch design comparisons reported for spoof detection.

2) Intra-Sensor, Cross-Material:

Cross-material experiments test whether the same-sensor detector generalizes to spoof materials absent from training. LivDet 2015 includes two unseen materials per sensor, and LivDet 2011 and 2013 are evaluated under a related protocol.

  • 2) Intra-Sensor, Cross-Material:: Cross-material testing uses the same sensor for training and testing, but introduces spoof materials not seen during training.This setting directly evaluates generalization to new fabrication materials.
  • 2) Intra-Sensor, Cross-Material:: LivDet 2015 cross-material tests use Liquid Ecoflex and RTV for three sensors, and OOMOO and Gelatin for Crossmatch.The proposed method is compared with state-of-the-art performance using Ferrfake unknown.
  • 2) Intra-Sensor, Cross-Material:: Cross-material performance is reported for LivDet 2011 and LivDet 2013 using average classification error and Ferrfake @ Ferrlive = 1%.The experiments follow the protocol adopted by the LivDet 2015 winner.

3) Cross-Sensor Evaluation:

The cross-sensor evaluation tests training and testing on different fingerprint readers within the same dataset, measuring whether the method learns distinctions transferable across acquisition devices.

  • Cross-Sensor Evaluation: Cross-sensor testing uses images from two different sensors within the same dataset.For LivDet 2011, Biometrika images are used for training and ItalData images for testing.
  • Cross-Sensor Evaluation: The evaluation measures the algorithm’s ability to learn characteristics distinguishing live and spoof fingerprints across acquisition devices.
  • Cross-Sensor Evaluation: ACE and Ferrfake @ Ferrlive = 1% are the reported metrics for cross-sensor comparison.

4) Cross-Dataset Evaluation:

The cross-dataset evaluation trains and tests on different datasets captured with the same sensor, assessing robustness to environmental changes in data collection. The proposed approach reduces average classification error but retains substantial false-accept error at the specified operating point.

  • Cross-Dataset Evaluation: Cross-dataset testing uses the same sensor but different datasets to capture invariance to changes in data-collection environments.One example trains on Biometrika images from LivDet 2011 and tests on Biometrika images from LivDet 2013.
  • Cross-Dataset Evaluation: 29% reduction in average classification error is reported, from 25.25% for to 17.91% for the proposed approach.
  • Cross-Dataset Evaluation: 52.52% and 65.06% average Ferrfake @ Ferrlive = 1% are reported for cross-sensor and cross-dataset scenarios, respectively.These results indicate challenges where a high level of spoof detection accuracy is needed.
  • Cross-Dataset Evaluation: The comparison is framed using average classification error and Ferrfake @ Ferrlive = 1%.

IV. FINGERPRINT SPOOF BUSTER

Fingerprint Spoof Buster is a graphical interface for real-time fingerprint spoof detection that lets operators select readers and trained CNN models, inspect color-coded local evidence, and review global decisions.

  • Fingerprint Spoof Buster: The interface supports real-time fingerprint spoof detection and selection of a fingerprint reader and trained MobileNet-v1 model.It offers online and batch evaluation modes.
  • Fingerprint Spoof Buster: Local fingerprint minutiae and patches are displayed with colors based on spoofness scores: green for live and red for spoof.
  • Fingerprint Spoof Buster: The interface is stated to be made available to interested researchers after publication acceptance.
  • Fingerprint Spoof Buster: Operators can visually examine highlighted local fingerprint regions alongside the global spoofness score and final decision.Batch mode outputs global spoofness scores for files in a specified directory.
  • Fingerprint Spoof Buster: The broader approach reports error-rate reductions of 63% intra-sensor, 43% cross-material, 4% cross-sensor, and 29% cross-dataset against state-of-the-art methods.
Loading 1712.04489v1…