EAGER: Exploring the Feasibility of Software Testing Techniques to Evaluate Fairness Algorithms in Software Systems

今天，软件正在做出更具社会影响力的自动化决策。例如，软件决定谁获得贷款或被雇用，计算风险评估分数，帮助决定谁入狱，谁被释放，并帮助诊断和治疗病人。软件在此类决策中的作用越来越大，使得软件公平性成为一个关键属性。随着越来越多的社会功能在网络空间中运作，软件公平性的重要性也在增加。本项目评估了使用软件测试技术来识别其输出对某些输入更有利的行为的可行性。以这种方式使用测试的目的是捕捉软件输入特性和软件行为之间的因果关系。 与分析数据但不测试应用软件行为的典型机器学习分类技术相比，该方法是新颖的。其中心思想是识别软件输入和软件行为方式之间的因果关系，例如，其产出。软件测试允许进行因果实验，包括使用几乎相同的输入运行软件，这些输入仅在测试中的关键特性上有所不同。影响行为的特征的变化提供了因果关系的证据。测量这样的因果关系需要测试套件，专注于在测试中的一组关键特性的小的可变性，而现有的测试技术专注于大的可变性，导致更大的覆盖范围。因此，现有的技术不适合测量因果关系，新技术是必要的。其结果是能够测试软件的新属性，没有测试程序存在。

Today, software is making more automated decisions with societal impact. For example, software determines who gets a loan or gets hired, computes risk-assessment scores that help decide who goes to jail and who is set free, and aids in diagnosing and treating medical patients. The increased role of software in such decisions makes software fairness a critical property. As more societal functions operate in cyberspace, the importance of software fairness increases. This project evaluates the feasibility of using software testing technology to identify behaviors whose outputs are more favorable for certain inputs. Using testing in such a manner is aimed at capturing causal relationships between characteristics of the software inputs and the software behavior. The approach is novel compared to typical machine learning classification techniques that analyze data but do not test the behavior of application software. The central idea is to identify causal relationships between software inputs and the way the software behaves, e.g., its outputs. Software testing enables conducting causal experiments consisting of running the software with nearly identical inputs that vary only in a key characteristic under test. Variations in that characteristic that affect behavior provide evidence of a causal relationship. Measuring such causal relationships requires test suites that focus on small variability in a key set of characteristics under test, while existing testing techniques focus on large variability that leads to greater coverage. As a result, existing techniques are ill-suited for measuring causal relationships and new technology is necessary. The result is the ability to test software for new properties for which no testing procedures existed.