SHF: Medium: Fairness in Software Systems

University Of Massachusetts Amherst, Amherst MA

Investigators

Abstract

Software impacts society in many ways and increasingly automates decision-making. For example, software transcribes videos, translates documents, selects what news articles are promoted, and determines who gets a loan or gets hired. It is possible for software to exhibit bias in its operation, whether or not it is intended by the customers or developers of the software. For example, software might be more accurate at transcribing male voices than female ones. Or software may inject societal stereotypes into automated translations, and risk-assessment computations may exhibit racial bias. As more societal functions operate in cyberspace, the importance of software fairness increases. In these settings, data-driven software has the ability to shape human behavior: it affects the products we view and purchase, the news articles we read, the social interactions we engage in, and, ultimately, the opinions we form. Biases in data and software risk forming, propagating, and perpetuating biases in society. This project develops theory, techniques and tools to enable software designers and engineers to describe fairness requirements, test the software for fairness properties, and debug fairness defects. The outcomes of this project will help increase the society's trust in software decisions and in the data the software uses, in turn, increasing potential impact and benefits the software can bring to society. The project addresses scientific questions behind efficiently and effectively measuring potential bias and helping stakeholders make informed decisions about software. It is not the project's aim to devise policies or eliminate bias in software. Instead, the aim is to provide software testing tools and measures that can be validated for formally specified software fairness properties. To measure bias, the project develops a novel approach for measuring causal relationships between program inputs and outputs. Software testing enables conducting causal experiments consisting of running the software with nearly identical inputs that vary only in a key input characteristic under test. Variations in an input characteristic that affect execution behavior provide evidence of a causal relationship. The project identifies when causal relationships are appropriate for measuring potential bias, develops efficient testing methods for measuring these relationships, and creates tools and techniques to help engineers identify and modify the causes of these relationships. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.