Work

Fairness and Ethical Analysis of Classification Models based on Facial Recognition

Year: 2023
Type: dissertation
Abstract: Facial recognition is a sub-field of AI that has seen tremendous growth over the last decade both in research and commercial applications. However, there has been growing criticism from scholars in various disciplines and media on the potential harms and biases that can be present in facial recognition technologies. Algorithmic fairness scholars have proposed the algorithmic audit as a means to uncover and mitigate harms of algorithmic systems. In this thesis, I audit fairness and ethical standards of classification algorithms that are based on facial recognition. Firstly, I investigate social classification algorithms that use facial recognition to predict a social class from someone’s face. I perform an audit on an algorithm that claims to predict political orientation from faces, and show that the algorithm shows unreliability when predicting on two images of the same person before and after some presentational change such as hair, makeup, glasses, or beard. Secondly, I audit facial emotion recognition (FER) algorithms that claim to predict a person’s emotion from an image of their face. I assess three popular FER datasets and three state-of-the-art models trained on each dataset. The results show that the both the datasets and models exhibit negative biases against people of darker skin color and people who are not observed as racially white. Lastly, I reflect on the process of auditing algorithms using the FER audit as a case study. I explore ways in which audits are justified and ways in which audits fall short at addressing affective presences arising from observing potential harms. I also explore how anger and hurtfulness motivated my criticisms of FER systems, and how the audits encourage suppression of such negative emotions. I further argue that the culture of affective regulation in algorithmic fairness research limits the potential for the field to effectively address harms of algorithmic systems.
Author: Khan, Rohan
Language: English
Open Access status: green
OpenAlex ID: W7056296711
Cites:
Cited by:
Related to: 9