Abstract
Growing cases of disparate outcomes due to Machine Learning (ML)-based systems have propelled notions of trustworthiness as a pressing concern rather than an afterthought. Although there is considerable debate about whether facial processing technology (FPT) fuels criminal justice disparities or can be used as a tool to ameliorate them, I investigate if the responsible design, development, evaluation and interpretation of FPT can help address racial inequalities in the Miami-Dade County (Florida, U.S.) criminal justice system.
Based on several interdisciplinary research questions, my research methodology proposes experimentation-based methods to address various fairness and bias issues within end-to-end deep learning image classification, I design an equitable methodology for generating and interpreting racial categories using mugshots from the Miami-Dade County criminal justice system. By considering race as a multidimensional construct, I assess the performance of eight deep CNN-based architectures when classifying mugshots according to binary race, and four race/ethnicity categories. By proposing a rigorous “self-auditing” model evaluation strategy, I provide empirical support for improving the disaggregated evaluation when predicting Black mugshots by 0.22% to 34.27% compared to White mugshots. Lastly, by implementing “post-hoc” gradient-based saliency maps, I assess the consistency of facial region relevance to a model when generating a racial prediction and make cautionary arguments for the use of a deep learning approach in a high-stake decision-making domain in an effort to foster greater ML trustworthiness within criminal justice stakeholders.