Evaluating Proposed Fairness Models for Face Recognition Algorithms

Abstract

The development of face recognition algorithms by academic and commercial organizations is growing rapidly due to the onset of deep learning and the widespread availability of training data. Though tests of face recognition algorithm performance indicate yearly performance gains, error rates for many of these systems differ based on the demographic composition of the test set. These "demographic differentials" in algorithm performance can contribute to unequal or unfair outcomes for certain groups of people, raising concerns with increased worldwide adoption of face recognition systems. Consequently, regulatory bodies in both the United States and Europe have proposed new rules requiring audits of biometric systems for "discriminatory impacts" (European Union Artificial Intelligence Act) and "fairness" (U.S. Federal Trade Commission). However, no international standard for measuring fairness in biometric systems yet exists. This paper characterizes two proposed measures of face recognition algorithm fairness (fairness measures) from scientists in the U.S. and Europe, using face recognition error rates disaggregated across race and gender from 126 commercial and open source face recognition algorithms. We find that both proposed methods have mathematical characteristics that make them challenging to interpret when applied to disaggregated face recognition error rates as they are commonly experienced in practice. To address this, we propose a set of interpretability criteria, termed the Functional Fairness Measure Criteria (FFMC), that outlines a set of properties desirable in a face recognition algorithm fairness measure. We further develop a new fairness measure, the Gini Aggregation Rate for Biometric Equitability (GARBE), and show how, in conjunction with the Pareto optimization, this measure can be used to select among alternative algorithms based on the accuracy/fairness trade-space. Finally, to facilitate the development of more robust and interpretable fairness measures in the face recognition domain, we have open-sourced our dataset of machine-readable, demographically disaggregated error rates. We believe this is currently the largest such dataset of its kind available to the ML fairness community.

On the Potential of Algorithm Fusion for Demographic Bias Mitigation in Face Recognition

“…The main drawback of FDR is that while its theoretical range of values is between 0 and 1, as in FFMC.2 required, it uses only a small portion of that range in practice. In fact, the range is mostly narrowed between 0:9 and 1, as shown in Howard et al's [73] study. Since 1 means fair and 0 means unfair, this fact could lead to the impression that all systems are fair, even if it is not the case.…”

“…While further discussion is required to standardize final definitions of fairness metrics, for this study, we follow the argumentation of Howard et al [73] and use GARBE to compare the fairness of different systems since it satisfies the most FFMCs. Additionally, GARBE can differentiate very well between fair and unfair compared to FDR and its fixed range is easier to interpret than the unbound IR.…”

Using biologically hierarchical modular architecture for explainable, tunable, generalizable, spatial AI

“…3,4 The seemingly omnipresent untunable black box problem in AI has been detrimental to the public perception and adoption of AI in a wide range of applications including LLMs, biometric recognition platforms, medical applications, and various mission-critical applications. [3][4][5] Of course, the black box problem is not exclusive to AI; analysis of black box algorithms is an implicit challenge in neuroscience. Connecting neuronal activity with behavior can be difficult because of the large variety of interpretations that are possible from neuronal activity alone (e.g., the past debate around view-based or object-centered object representation 6 ).…”