Title:	BUILDING A FACE DATABASE OF ARAB FACES TOWARD EVALUATING BIAS IN FACIAL ANALYSIS SYSTEMS
Author(s):	Ashraf Khalil, Suha Glal, Khider Ahmed, Sana Zeb Khan and Aysha Abdulgani
ISBN:	978-989-8704-49-8
Editors:	Katherine Blashki, Yingcai Xiao, Piet Kommers and Pedro Isaías
Year:	2023
Edition:	Single
Keywords:	Arab Face Database, Arab Public Figures Face, Algorithmic Audit, Image Manipulation, Image Processing, Machine Learning
Type:	Full
First Page:	63
Last Page:	70
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	Machine learning algorithms are fundamentally driven by the data provided by humans consequently, the decisions made by those algorithms are not free from human bias. This is particularly evident in the case of facial analysis systems that employ machine learning algorithms. Recent studies have shown that the decisions made by many of the commercially available facial analysis systems are prejudiced against certain groups of race, ethnicity, age, gender and culture. Further studies have identified that the underlying reason for such biased decisions is that the open source material available for facial image databases which are used in commerce and academia to train the algorithms has meager diversity in these categories. To compound this issue, facial analysis technology is promoted by influential companies and artificial intelligence service providers without affirming the fairness and accuracy of the decisions given by these systems. To minimize bias and ensure representation of the Middle Eastern population in the imminent growth of this technology, we propose the development of two Arab face databases along with an algorithmic audit involving seven commercially available facial analysis systems. Of the databases, the first, Arab-LEANA, will include 300 Arab subjects' face images with variation in lighting, expression, accessory, nationality and age (LEANA). The second, Arab Public Figures Faces (APFF), will contain images and videos of 300 Arab public figures captured "in the wild". Faces for APFF will be selected manually from the internet since manual selection of faces will result in a high degree of variability in scale, pose, expression, illumination, age, occlusion and make-up. These databases will provide the worldwide community of face recognition researchers with a large-scale, diverse collection of Arab face images for training and evaluating algorithms toward developing a more representative, and therefore more robust, capacity for facial analysis. This, in turn, will facilitate the development of more accurate face recognition technology as it prepares to go mainstream and enter numerous facets of modern life.
	Go Back