Source-linked AI summary
A high performance fingerprint liveness detection method based on quality related features
Javier Galbally, Fernando Alonso-Fernandez, Julian Fierrez, Javier Ortega-Garcia
TL;DR
Fingerprint liveness detection must address direct attacks in which synthetic traits are presented to sensors and recognized as genuine. The paper proposes a software-based classifier using ten quality-related fingerprint features, with sensor-dependent feature selection and a single acquired image. Across diverse public datasets and attack scenarios, the system correctly classifies almost 90% of fingerprint images while reducing acquisition burden and hardware requirements.
Problem
Biometric systems are vulnerable to direct attacks using synthetic traits, motivating countermeasures that distinguish fake samples from genuine ones.
Method
A software-based liveness system extracts ten quality-related features from one fingerprint image and selects features according to the acquisition sensor.
Results
Almost 90% of fingerprint images were correctly classified across diverse sensors, gummy-finger materials, and generation procedures.
Takeaways & Limitations
Using one acquired image makes the approach less intrusive, faster, more user friendly, and lower cost than approaches requiring additional hardware or acquisition steps.
Takeaways & Limitations
Feature selection depends on the acquisition sensor, while the known fake-finger material is not used by the system.
Abstract
from arXiv · showhide
A new software-based liveness detection approach using a novel fingerprint parameterization based on quality related features is proposed. The system is tested on a highly challenging database comprising over 10,500 real and fake images acquired with five sensors of different technologies and covering a wide range of direct attack scenarios in terms of materials and procedures followed to generate the gummy fingers. The proposed solution proves to be robust to the multi-scenario dataset, and presents an overall rate of 90% correctly classified samples. Furthermore, the liveness detection method presented has the added advantage over previously studied techniques of needing just one image from a finger to decide whether it is real or fake. This last characteristic provides the method with very valuable features as it makes it less intrusive, more user friendly, faster and reduces its implementation costs.
1. Introduction
Biometric systems face vulnerabilities because biometric traits are exposed, irreplaceable, and susceptible to direct attacks using synthetic traits. The paper proposes a software-based fingerprint liveness method using quality-related features, evaluated across diverse attack scenarios.
- Biometric systems are vulnerable because fingerprints can be obtained and cannot be replaced if stolen.
- Direct attacks present synthetic fingerprints to sensors so they are recognized as legitimate users and access is granted.
- Successful direct attacks require recovering a legitimate biometric trait, generating an artefact, and having the system recognize it as genuine.
- Liveness detection uses physiological properties to distinguish real from fake traits while seeking non-invasive, user-friendly, fast, and low-cost operation.
- The proposed method uses quality-related measures and one fingerprint image to detect fake fingers, shortening acquisition and reducing user inconvenience.
- Almost 90% of images were correctly classified on a dataset exceeding 10,500 real and fake images from five sensors and varied gummy-finger scenarios.
2. Related Works
Fingerprint liveness research includes software- and hardware-based techniques that measure perspiration, elasticity, odor, electrical properties, texture, or spectral characteristics. Software-based methods operate on acquired images and avoid the extra cost and intrusiveness of added hardware.
- Liveness detection approaches have been developed for fingerprints, faces, and irises.
- Software-based techniques: Software-based techniques detect fake traits from samples acquired with standard sensors rather than measuring the finger with added hardware.
- Hardware-based techniques: Hardware-based techniques add devices to measure living-trait properties such as blood pressure, odor, or pupil hippus.
- Software-based solutions are less expensive and less intrusive because they require no additional device.
- Fingerprint methods have examined perspiration, elasticity, odor, electrical properties, wavelets, power spectra, and Fourier-spectrum ring patterns.
- LivDet compared software-based fingerprint liveness solutions and estimated effective static and dynamic features.
3. Liveness Detection System
The proposed system treats liveness detection as binary classification using a sensor-aware vector of ten fingerprint image quality features. It segments the fingerprint, extracts quality measures, selects features by sensor, and classifies the image as real or fake.
- Liveness detection is formulated as assigning an input fingerprint image to either the real or fake class.
- The system takes a fingerprint image and the acquisition sensor as inputs.
- Fingerprint segmentation uses Gabor filters, after which ten quality measures form the feature vector for classification.
- Quality assessment measures ridge strength or directionality, ridge continuity, ridge clarity, ridge-valley integrity, or estimated verification performance.
- Ridge-strength measures: Orientation Certainty Level measures energy concentration along dominant ridge directions using the gradient covariance eigenvalue ratio.
- Ridge-strength measures: Power-spectrum energy concentration uses ring-shaped frequency bands and entropy, with concentrated energy indicating higher quality.
- Ridge-strength measures: Local Orientation Quality averages directional differences with surrounding blocks, while orientation-field continuity accumulates abrupt changes.
3.2. Feature Selection
Feature selection evaluates all possible subsets of the ten proposed features, with the optimal subset potentially differing by acquisition sensor. LDA is trained and evaluated using separate development and test sets.
- Feature subset search: 1,023 feature subsets are exhaustively searched to identify the best-performing feature subset.The search covers all possible subsets of the ten proposed features.
- Sensor dependence: The optimal feature subset may depend on the sensor used for acquisition.Feature selection is performed separately according to the acquisition device.
- Classification: Linear Discriminant Analysis is used to classify samples as real or fake.The classifier is fitted using development data after feature selection.
- Experimental split: Development and test sets are kept separate for feature selection, model fitting, and performance evaluation.The development set selects features and fits class distributions, while the test set evaluates the algorithm.
4. Datasets and Experimental Protocol
The protocol evaluates liveness detection across LivDet 2009 and ATVS databases spanning multiple sensors, fake-finger materials, and generation procedures. The final dataset contains over 10,500 real and fake samples and the system uses only the image and acquisition sensor as inputs.
- LivDet Database: LivDet includes three flat optical sensors and gummy fingers made from silicone, gelatine, and playdoh using a consensual procedure.The sensors are Biometrika FX2000, CrossMatch Verifier 300CL, and Identix DFR2100.
- ATVS Database: ATVS includes optical, capacitive, and thermal sweeping devices, with silicone gummy fingers generated with and without user cooperation.The devices are Biometrika FX2000, Precise SC100, and Yubee with Atmel’s Fingerchip.
- Dataset composition: Over 10,500 real and fake samples form the final validation dataset across six datasets.Three datasets come from LivDet and three from ATVS.
- Scenario variation: The datasets vary by acquisition device, material used to generate gummy fingers, and procedure used to obtain fake images.These factors create different direct-attack scenarios.
- System inputs: The system receives only the fingerprint image and the sensor used for acquisition.It does not use knowledge of the fake-finger material, and feature selection depends only on the sensor.
- Problem difficulty: The final classification problem is difficult enough that distinguishing real from fake samples would also challenge a human expert.This difficulty is illustrated by examples in Figures 8 and 9.
5. Results
Performance is estimated using Average Classification Error, which combines the false liveness and false fake rates. The supplied passage provides the metric definition but no numerical result.
- Evaluation metric: Average Classification Error is the performance measure used for the proposed liveness detection scheme.It is defined as ACE = (FLR + FFR)/2.
- Evaluation metric: ACE averages the false liveness rate and false fake rate.The formula explicitly combines FLR and FFR equally.
- Reported result: The supplied results passage defines ACE but does not report a numerical ACE value.
BIOMETRIKA FX2000 (FLAT OPTICAL SENSOR)
The supplied passage contains only the label “Silicone” and provides no substantive experimental result or description for this section.
- BIOMETRIKA FX2000 (FLAT OPTICAL SENSOR): The passage identifies silicone.
- BIOMETRIKA FX2000 (FLAT OPTICAL SENSOR): The passage repeats the material label without adding experimental details.
- BIOMETRIKA FX2000 (FLAT OPTICAL SENSOR): No sensor-specific performance value is provided in the passage.
FAKE (WITH USER COOPERATION)
The validation evaluates quality-based liveness detection across diverse sensors, materials, datasets, and attack-generation procedures. Performance is generally strong, although the LivDet Biometrika dataset shows a pronounced training–validation discrepancy.
- Feature selection: Minimum ACE occurred with subsets containing 6 or 7 parameters, while adding further features slightly increased error.This pattern was observed across all six datasets and reflects a curse-of-dimensionality effect.
- Feature selection: All proposed quality features appeared in at least one optimal subset, while QE and QOCL occurred in every dataset’s best subset.The authors associate the most discriminant features with ridge strength, alongside selected ridge clarity and continuity measures.
- Feature selection: The Biometrika FX2000 produced the same best parameterization in the LivDet and ATVS datasets.This suggests consistency of the selected feature subsets across different scenarios when the acquisition device is unchanged.
- Validation: The overall classification error was around 10% in the cross-validation evaluation.The reported ACE combines the two cross-validation stages by averaging ACE1 and ACE2.
- Validation: The LivDet Biometrika dataset showed ACE increasing from 1.73% during training to 26.5% during validation.The discrepancy included an abnormally high False Fake Rate above 50% when the development set trained the classifier and the test set evaluated it.
- Validation: The authors estimate a more realistic ACE of around 10% for Biometrika.LD and an overall system error of around 7.5% under that assumption.This estimate is intended to account for the suspected inconsistency between the LivDet development and test data.
- FAKE (WITH USER COOPERATION): Non-cooperative attacks had lower ACE than cooperative attacks for every studied ATVS case.The authors link this to the more difficult generation process and the resulting greater separation of non-cooperative quality distributions from genuine images.
6. Conclusions
The paper proposes quality-based fingerprint parameterization for software liveness detection and evaluates it across diverse public datasets and attack scenarios. It reports nearly 90% correct classification while using only the acquired fingerprint image, making acquisition faster and less intrusive.
- 6. Conclusions: The proposed parameterization uses quality-related measures to detect fake fingerprints in software-based liveness detection.It was evaluated on LivDet 2009 and the ATVS database across different sensor technologies, gummy-finger materials, and generation procedures.
- 6. Conclusions: Almost 90% of fingerprint images were correctly classified under diverse testing scenarios.The authors present this result as evidence of performance across different direct-attack conditions.
- 6. Conclusions: The method classifies each acquired fingerprint image using software features rather than additional physiological measurements from special hardware.The approach therefore avoids hardware additions associated with higher sensor cost and greater user intrusiveness.
- 6. Conclusions: Using features extracted from one image makes acquisition faster and more convenient for the user.The user need not keep the finger on the sensor for several seconds or place it multiple times.
- 6. Conclusions: Liveness detection helps prevent direct attacks using synthetic traits and enhances the security offered by fingerprint verification systems.The conclusion frames this as the practical importance of the presented class of solutions.