mdfa: Multi-Differential Fairness Auditor for Black Box Classifiers

Abstract

Machine learning algorithms are increasingly involved in sensitive decision-making process with adversarial implications on individuals. This paper presents mdfa, an approach that identifies the characteristics of the victims of a classifier's discrimination. We measure discrimination as a violation of multi-differential fairness. Multi-differential fairness is a guarantee that a black box classifier's outcomes do not leak information on the sensitive attributes of a small group of individuals. We reduce the problem of identifying worst-case violations to matching distributions and predicting where sensitive attributes and classifier's outcomes coincide. We apply mdfa to a recidivism risk assessment classifier and demonstrate that individuals identified as African-American with little criminal history are three-times more likely to be considered at high risk of violent recidivism than similar individuals but not African-American.

Bias and unfairness in machine learning models: a systematic literature review

“…The work allows for black-box type algorithms to be audited for bias mitigating injustice, but also understanding the treatment. It further reinforces that injustices can be identified by performing a comparison between the treatment received by one group versus another, through the MDFA, which aims to ensure that minority groups receive proper treatment regardless of their sensitive attributes [24].…”

“…Except for [24] and [8], the model proposals were primarily white-box classification. The former proposes a model for bias elimination using Multi-Differential Fairness by integrating in-processing and post-processing, whereas the latter proposes that the focus of algorithm transparency should be on the output rather than the whole decision-making process of the algorithm.…”

“…In addition, the authors argue that it is possible to increase the quality of the technique by combining it with others, for example, using a fairer base than the one used in the experiments. [24] proposes a solution with the post-processing method of an already trained model, seeking to identify whether certain groups receive a discriminatory treatment due to their sensitive attributes. With the identification of discrimination for a group, it is verified whether or not sensitive attributes are being passed to the model, even if indirectly.…”

“…Both [23] and [24] present a post-processing model for identifying the group most disadvantaged by the decision algorithm, but for mitigation and identification, respectively. [23] emphasizes that there are numerous tools for postprocessing, such as AEQUITAS, and techniques such as Adversarial Learning, which is also highlighted in [22].…”

“…In the [24] study, the model receives as input the attributes of the original dataset, whether their are sensitive or not, together with the output of the algorithm being evaluated. After that, it analyzes the MDFA, as the measure of disparity in the treatment received by groups with different sensitive attributes.…”