Publicación: Measuring Bias and Improving Fairness and Accuracy of Deep Face Representations
Cargando...
Fecha
2020-06-01
Autores
Editor/a
Director/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Nacional de Educación a Distancia (España). Escuela Técnica Superior de Ingeniería Informática. Departamento de Inteligencia Artificial
Resumen
Este trabajo explora los sesgos en los procesos de aprendizaje basados en arquitecturas de redes neuronales profundas mediante el estudio de la biometría facial. Los indicadores de referencia más populares para el reconocimiento facial asumen una distribución de los sujetos sin prestar mucha atención a sus atributos demográficos. En este trabajo se muestra experimentalmente que los procesos de aprendizaje basados en las bases de datos de cara más utilizadas han dado lugar a populares modelos de redes neuronales que presentan una fuerte discriminación algorítmica. En el presente trabajo hacemos un análisis detallado de las causas y efectos del sesgo en las características aprendidas por las redes neuronales profundas utilizadas para la biometría facial, con un amplio estudio demográfico de las bases de datos más utilizadas para entrenar estos modelos. También mostramos cómo los rasgos étnicos influyen en las activaciones de los modelos de detección de género basados en imágenes de cara. Los experimentos incluyen dos modelos populares de reconocimiento facial y tres bases de datos públicas compuestas por 64.000 identidades de diferentes grupos demográficos distinguidos por el género y la etnia. Proporcionamos una formulación general de la discriminación algorítmica. Finalmente proponemos InsideBias, un novedoso método para detectar modelos sesgados, y SensitiveLoss, un nuevo método de aprendizaje que tiene en cuenta la discriminación para mejorar tanto la precisión como la imparcialidad de los algoritmos de reconocimiento facial. InsideBias se basa en la forma en que los modelos representan la información en lugar de cómo rinden, que es la práctica habitual en otros métodos existentes para la detección de sesgos. Nuestra estrategia con InsideBias permite detectar modelos sesgados con muy pocas muestras (sólo 15 imágenes en nuestro estudio de caso). SensitiveLoss se basa en la popular función de pérdida triplet loss y en un generador de tripletes sensibles. Este enfoque funciona como un complemento de las redes pre-entrenadas y se utiliza para mejorar su rendimiento en términos de precisión e imparcialidad. El método muestra resultados comparables a los de las redes sin sesgo del estado del arte y representa un paso hacia delante para evitar los efectos discriminatorios de los sistemas automáticos.
This work explores the biases in learning processes based on deep neural network architectures through the study of face biometrics. The most popular face recognition benchmarks assume a distribution of subjects without paying much attention to their demographic attributes. In this work, we experimentally show that learning processes based on the most used face databases have led to popular pre-trained deep face models that present a strong algorithmic discrimination. We make a detailed analysis of the causes and effects of bias on the features learned by deep neural networks used for face biometrics, with an extensive demographic study of the databases used to train these models. We also show how ethnic attributes impact in the activations of gender detection models based on face images. The experiments include two popular face recognition models and three public databases composed of 64,000 identities from different demographic groups characterized by gender and ethnicity. We provide a general formulation of algorithmic discrimination. We finally propose InsideBias, a novel method to detect biased models, and SensitiveLoss, a new discrimination-aware learning method to improve both accuracy and fairness of face recognition algorithms. InsideBias is based on how the models represent the information instead of how they perform, which is the normal practice in other existing methods for bias detection. Our strategy with InsideBias allows to detect biased models with very few samples (only 15 images in our case study). SensitiveLoss is based on the popular triplet loss function and a sensitive triplet generator. This approach works as an add-on to pre-trained networks and is used to improve their performance in terms of average accuracy and fairness. The method shows results comparable to state-of-the-art debiasing networks and represents a step forward to prevent discriminatory effects by automatic systems.
This work explores the biases in learning processes based on deep neural network architectures through the study of face biometrics. The most popular face recognition benchmarks assume a distribution of subjects without paying much attention to their demographic attributes. In this work, we experimentally show that learning processes based on the most used face databases have led to popular pre-trained deep face models that present a strong algorithmic discrimination. We make a detailed analysis of the causes and effects of bias on the features learned by deep neural networks used for face biometrics, with an extensive demographic study of the databases used to train these models. We also show how ethnic attributes impact in the activations of gender detection models based on face images. The experiments include two popular face recognition models and three public databases composed of 64,000 identities from different demographic groups characterized by gender and ethnicity. We provide a general formulation of algorithmic discrimination. We finally propose InsideBias, a novel method to detect biased models, and SensitiveLoss, a new discrimination-aware learning method to improve both accuracy and fairness of face recognition algorithms. InsideBias is based on how the models represent the information instead of how they perform, which is the normal practice in other existing methods for bias detection. Our strategy with InsideBias allows to detect biased models with very few samples (only 15 images in our case study). SensitiveLoss is based on the popular triplet loss function and a sensitive triplet generator. This approach works as an add-on to pre-trained networks and is used to improve their performance in terms of average accuracy and fairness. The method shows results comparable to state-of-the-art debiasing networks and represents a step forward to prevent discriminatory effects by automatic systems.
Descripción
Categorías UNESCO
Palabras clave
aprendizaje profundo, Imparcialidad, sesgo algorítmico, IA explicable, comportamiento máquina, discriminación, aprendizaje automático, cara, biometría, deep e-Learning, fairness, bias, explainable ai, machine behavior, discrimination
Citación
Centro
E.T.S. de Ingeniería Informática
Departamento
Inteligencia Artificial