Publication: Estudio del rendimiento biométrico de dispositivos de huella dactilar : análisis de calidad y variabilidad de la muestra
Loading...
Identifiers
Publication date
2015
Defense date
2015-10-16
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
El trabajo presentado se centra en el estudio de rendimiento de sensores biométricos basados en huella dactilar. El estudio concretamente analiza la influencia de la calidad de la huella. Para poder determinar la calidad de estas huellas se ha usado el algoritmo NFIQ, este algoritmo asigna a cada muestra biométrica un valor de calidad. Estos valores se han distribuido entre tres niveles de calidad (alto, medio y bajo), dividiendo de esta forma las muestras. Finalmente, se han desarrollado tres estudios de rendimiento, uno por cada nivel de calidad, y se han comparado los resultados.
Para poder ver que la influencia de NFIQ es equivalente en cualquier sensor biométrico, el estudio ha sido llevado con tres sensores de tres fabricantes distintos. Además, para poder tener suficientes valores se ha pedido la colaboración de 50 usuarios, obteniendo más de 6000 muestras.
Una vez obtenidas las muestras, éstas han sido procesadas haciendo comparaciones entre ellas. Estas comparaciones permiten conocer el rendimiento de los sensores para los diferentes niveles de calidad en cada sensor. Estas comparaciones se dividen en dos grupos ya sean comparaciones entre imágenes de un mismo usuario o entre usuarios distintos. Las llamadas comparaciones genuinas y de impostores, respectivamente.
Al obtener estas listas se ha podido representar una serie de valores que indican el porcentaje de usuarios que conseguirían vulnerar el sistema, el valor mínimo de comparación para hacer el sistema más seguro a cambio de ser más restrictivo, etc.
Finalmente, al analizar estas gráficas se observó cómo los valores NFIQ de menor calidad tenían una clara influencia negativa sobre el rendimiento de los sensores; pero los valores de calidad media y alta mostraban resultados muy similares. Es por esto que se plantea nuevos estudios que permitan un mayor análisis del algoritmo NFIQ para discernir si es necesario la mejora del algoritmo para notar mayor diferencia entre las calidades.
The former study is intended to bring a new point of view on biometric identification. It focuses on studying how valid the approximation value of the algorithm NFIQ, made by the NIST, has on the actual overall quality of a biometric system. The project will use 3 different fingerprint sensors, these sensors are from a different manufacturer each. The purpose of these sensors is to see if the use of different technology has any influence on the NFIQ algorithm, moreover thanks to the sensors the study can assure that the results obtained would not only apply to a specific product but for a great variety ones. To carry out the research 50 users have been ask to assist by interacting with the three sensors. Each user generated at least 3 Enrol patrons with each finger, from thumb to the middle finger, in addition the users will go through the process of verification 6 times from each fingers. Both processes was carried out for each sensor. Once all the images were taken, these would be processes with a C# program. The process consisted on creating a list of genuine comparison values and impostor comparison values for each sensor. To obtain the first list, each finger would be assign a enrol patron, the one with the lowest NFIQ value (therefore the highest quality), and compare that patron with all the verifications images. The set of numbers of all the genuine comparison of each finger would become the final list of genuine comparison of the sensor those images belonged to. To obtain the second list, each finger Enrol would be compared to all the verification images of all other users. This las list was significantly bigger since there were more comparisons to be made. Once again, the set of values would be grouped together and presented as the impostor comparison values of the specific sensor. This process would be carry out for each quality level. After recollected the data that has been created, the different graphs needed to see the results that had been obtained were represented. When searching for the performance of a biometric sensor it is important to check 3 different graphs: the DET, which represents the FRR in dependence of the FAR (in percentage); the ROC, which shows for a percentage of admitted user how many would be impostor; and finally the FARvsFRR, which is informs of how the FRR and FAR vary when increasing a threshold value of comparison. For each sensor there would be 1 of these graphs with the values of low, medium and high quality on the same graph. Finally to be able to compare the sensors there would be an extra three groups of graphs comparing the different sensors at different quality levels. Thanks to this graphs, some conclusion were able to be obtained. As expected the NFIQs 5-4 were very much divided with other qualities. The values obtained at this level so very poor quality to the point of being of no use in a real life system. Nevertheless, the NFIQs of value 3 were very much a surprise due to the great results obtained. These values were comparable to those of value 1-2 even to the point of matching them in some cases. On the final conclusions, it was reached that the NFIQ algorithm still has place for improvement. The difference between the low quality values and the medium quality were too high in comparison with the difference with medium and high. It could be of interest to try to improve the NFIQ algorithm so that these values are better distributed. As a conclusion dedicated for the future companies that want to include biometric recognition as part of their security system, it was suggested to encourage a study on what are the exact influences on the NFIQ values. This means, what environmental or biological characteristics makes one fingerprint better than the other. If the factors that lower the quality of the images can be taken out, there will be more warranties of using these sensor in a security system.
The former study is intended to bring a new point of view on biometric identification. It focuses on studying how valid the approximation value of the algorithm NFIQ, made by the NIST, has on the actual overall quality of a biometric system. The project will use 3 different fingerprint sensors, these sensors are from a different manufacturer each. The purpose of these sensors is to see if the use of different technology has any influence on the NFIQ algorithm, moreover thanks to the sensors the study can assure that the results obtained would not only apply to a specific product but for a great variety ones. To carry out the research 50 users have been ask to assist by interacting with the three sensors. Each user generated at least 3 Enrol patrons with each finger, from thumb to the middle finger, in addition the users will go through the process of verification 6 times from each fingers. Both processes was carried out for each sensor. Once all the images were taken, these would be processes with a C# program. The process consisted on creating a list of genuine comparison values and impostor comparison values for each sensor. To obtain the first list, each finger would be assign a enrol patron, the one with the lowest NFIQ value (therefore the highest quality), and compare that patron with all the verifications images. The set of numbers of all the genuine comparison of each finger would become the final list of genuine comparison of the sensor those images belonged to. To obtain the second list, each finger Enrol would be compared to all the verification images of all other users. This las list was significantly bigger since there were more comparisons to be made. Once again, the set of values would be grouped together and presented as the impostor comparison values of the specific sensor. This process would be carry out for each quality level. After recollected the data that has been created, the different graphs needed to see the results that had been obtained were represented. When searching for the performance of a biometric sensor it is important to check 3 different graphs: the DET, which represents the FRR in dependence of the FAR (in percentage); the ROC, which shows for a percentage of admitted user how many would be impostor; and finally the FARvsFRR, which is informs of how the FRR and FAR vary when increasing a threshold value of comparison. For each sensor there would be 1 of these graphs with the values of low, medium and high quality on the same graph. Finally to be able to compare the sensors there would be an extra three groups of graphs comparing the different sensors at different quality levels. Thanks to this graphs, some conclusion were able to be obtained. As expected the NFIQs 5-4 were very much divided with other qualities. The values obtained at this level so very poor quality to the point of being of no use in a real life system. Nevertheless, the NFIQs of value 3 were very much a surprise due to the great results obtained. These values were comparable to those of value 1-2 even to the point of matching them in some cases. On the final conclusions, it was reached that the NFIQ algorithm still has place for improvement. The difference between the low quality values and the medium quality were too high in comparison with the difference with medium and high. It could be of interest to try to improve the NFIQ algorithm so that these values are better distributed. As a conclusion dedicated for the future companies that want to include biometric recognition as part of their security system, it was suggested to encourage a study on what are the exact influences on the NFIQ values. This means, what environmental or biological characteristics makes one fingerprint better than the other. If the factors that lower the quality of the images can be taken out, there will be more warranties of using these sensor in a security system.
Description
Keywords
Biometría, Huella dactilar, Sensores, Procesamiento de datos, Algoritmos, Calidad, Biometric, Fingerprint recognition, Sensors, Data processing, Algorithms, Quality control