Inclusive representation for Face Recognition analysis: an empirical investigation on critical issues and uncovered opportunities

Face recognition technology is frequently used, in modern applications, in different fields: security, advertising, autonomous car, banks and education. A variety of recent works show the threats and limitations of this technology, regarding vulnerable subjects: black population, in particular black females, young population, LGBTQ+ community and many others. Face features (e.g. lips, cheeks, eyes) and their measurements are the most important features of face recognition algorithms. In this context, this study aims at finding if there are any limitations of the technology itself applied to a population characterized by slightly different facial features, Down syndrome people. Due to the high sensitivity of the subjects, few works have been proposed in the literature, and they are mostly concerned on the recognition of the Syndrome. The current study includes the creation of a specific dataset, which was strictly related to the lack of existent resources of a dataset including Down syndrome population. The dataset, created for the purpose of the analysis, is composed of two different groups of images, the experimental group, representing Down syndrome people, and the control group, representing people without any syndrome. The core of the current study consists on testing two commercial algorithms of face recognition, AWS Rekognition and ClarifAI. The results obtained are characterized by three following important considerations. The gender recognition results show an high error rate towards the category of Down syndrome male. In particular, both algorithms predict in a correct way 85\% of images of the experimental group male class, in contrast with 97% and 94% of the control group male class for ClarifAI and AWS respectively. Moreover, a deeper analysis of the confidence values of each prediction, shows that only 66% and 60% of the images belonging to the experimental group male class, have a confidence value greater than or equal to 99.0, in contrast with 90% and 75% of the images of the control group male class. Regarding the results of the age prediction, 6% of the images representing people between 20 and 39 years old, belonging to the experimental group, are classified with 3-9 age range and 9% of the images representing people between 30 and 49 years old, belonging to the experimental group, are classified with 10-19 age range by ClarifAI algorithm. Instead, regarding the prediction of AWS algorithm, 14% of images representing people with true age range between 30 and 39 years old, belonging to the experimental group, are classified with 3-9 or 10-19 age range by AWS algorithm. These results lead to the conclusion that both algorithms assign children age range to adults belonging to the experimental group. The labels used by the models, to classify the images, are very different between each others. The AWS labels are more descriptive, instead the ClarifAI labels contains some adjectives and subjective considerations about the people in the images. The results obtained by those algorithms do not show significant differences between the experimental group and the control group, but they show very important differences between genders. In particular, labels of the Aesthetics category are more associated to female classes rather than male classes and labels of the Education category are more associated to male classes rather than female classes. In conclusion, the prediction of gender and age show meaningful differences between experimental and control group.