Fairness testing: testing software for discrimination

“nition of discrimination is that a “better” input is never deprived of the “better” output [24].”

“Our causal discrimination score is a stronger measure of discrimination, as it seeks out causality in software, measuring the fraction of inputs for which changing specific input characteristics causes the output to change [67].”

“While prior work has tackled the oracle problem [10, 26, 30] typically using inferred pre- and post-conditions or documentation, our oracle is more precise and easier to compute, but is only applicable to fairness testing.”

“Apparent discrimination captures this notion, applying group or causal discrimination score measurement to a subset of the input domain, which can be described by an operational profile [7, 58].”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“This dataset is well vetted: it has been used by others to devise discriminationfree algorithms [18, 88, 89], as well as for non-discrimination purposes [3, 43, 86].”

“, advertisements [75], hotel bookings [53], and image search [45].”

“Apparent discrimination captures this notion, applying group or causal discrimination score measurement to a subset of the input domain, which can be described by an operational profile [7, 58].”

“The Calders-Verwer (CV) score [19] measures the strength of group discrimination as the difference between the largest and the smallest outcome fractions; if 30% of people <40 get the loan, and 40% of people >40 get the loan, then loan is 40% − 30% = 10% group discriminating.”

“AdFisher computes a variant of group discrimination, but it could be integrated with Themis and its algorithms to measure causal discrimination.”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“, advertisements [75], hotel bookings [53], and image search [45].”

“, [18, 40, 88, 91]) produce fair algorithms? RQ2: How effective do the optimizations from Section 4 make Themis at identifying discrimination in software?”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“, [18, 40, 88, 91]) produce fair algorithms? RQ2: How effective do the optimizations from Section 4 make Themis at identifying discrimination in software?”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“This dataset is well vetted: it has been used by others to devise discriminationfree algorithms [18, 88, 89], as well as for non-discrimination purposes [3, 43, 86].”

“Our causal fairness definition is designed specifically to be testable, unlike definitions that require probabilistic estimation or knowledge of the future [38].”

“FairTest, an implementation of the unwarranted associations framework [79] uses manually written tests to measure four kinds of discrimination scores: the CV score and a related ratio, mutual information, Pearson correlation, and a regression between the output and sensitive inputs.”

“For example, all-pairs testing generates tests that evaluate every possible value combination for every pair of input characteristics, which can be particularly helpful when testing software product lines [6, 44, 46, 47].”

“Already, software makes decisions in what products we are led to buy [53], who gets a loan [62], self-driving car actions that may lead to property damage or human injury [32], medical diagnosis and treatment [74], and every stage of the criminal justice system including arraignment and sentencing that determine who g…”

“D is amodified decision tree approach that balances the training dataset to equate the number of inputs that lead to each output value, removes and repeats some inputs, and flips output values of inputs close to the decision tree’s decision boundaries, introducing noise around the critical boundaries [91].”

“While prior work has tackled the oracle problem [10, 26, 30] typically using inferred pre- and post-conditions or documentation, our oracle is more precise and easier to compute, but is only applicable to fairness testing.”

“This dataset is also well vetted: it has been used by others to devise discrimination-free algorithms [18, 89], as well as for non-discrimination purposes [25, 31].”

“, using static or dynamic analyses for vacuity testing [8, 33] or to identify configuration options that cannot affect a test’s output [48], can directly improve efficiency of discrimination testing by identifying that changing a particular input characteristic cannot affect a particular test’s output, and thus no cau…”

“For example, all-pairs testing generates tests that evaluate every possible value combination for every pair of input characteristics, which can be particularly helpful when testing software product lines [6, 44, 46, 47].”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“Causal relationships in data management systems [54, 55] can help explain query results [57] and debug errors [82–84] by tracking and using data provenance [56].”

“While prior work has tackled the oracle problem [10, 26, 30] typically using inferred pre- and post-conditions or documentation, our oracle is more precise and easier to compute, but is only applicable to fairness testing.”

“The goal of the generated tests is typically finding bugs [29] or generating specifications [21].”

“The number of tests needed to evaluate every possible value pair can be significantly smaller than the exhaustive testing alternative since each test can simultaneously contribute to multiple value pairs [22, 44, 80].”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“For example, all-pairs testing generates tests that evaluate every possible value combination for every pair of input characteristics, which can be particularly helpful when testing software product lines [6, 44, 46, 47].”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“Restricting the input space to real-world inputs [1] may similarly hide software discrimination that causal testing can reveal.”

“, as a result of implementation bugs, as an unintended property of self-organizing systems [11, 13, 15, 16], as an emergent property of component interaction [12, 14, 17, 49], or as an automatically learned property from biased data [18, 19, 39–42, 88, 89, 91].”

“, as a result of implementation bugs, as an unintended property of self-organizing systems [11, 13, 15, 16], as an emergent property of component interaction [12, 14, 17, 49], or as an automatically learned property from biased data [18, 19, 39–42, 88, 89, 91].”

“, as a result of implementation bugs, as an unintended property of self-organizing systems [11, 13, 15, 16], as an emergent property of component interaction [12, 14, 17, 49], or as an automatically learned property from biased data [18, 19, 39–42, 88, 89, 91].”

“Our prior work on testing software that relies on data management systems has focused on data errors [59, 60], whereas this work focuses on testing fairness.”

“, as a result of implementation bugs, as an unintended property of self-organizing systems [11, 13, 15, 16], as an emergent property of component interaction [12, 14, 17, 49], or as an automatically learned property from biased data [18, 19, 39–42, 88, 89, 91].”

“Our prior work on testing software that relies on data management systems has focused on data errors [59, 60], whereas this work focuses on testing fairness.”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“Because developer-written test suites are often not as representative of field executions as developers would like [81], operational profiles can significantly improve the effectiveness of testing by more accurately representing real-world system use [7, 50] and the use of operational profiles has been shown to more a…”

“The number of tests needed to evaluate every possible value pair can be significantly smaller than the exhaustive testing alternative since each test can simultaneously contribute to multiple value pairs [22, 44, 80].”

“, as a result of implementation bugs, as an unintended property of self-organizing systems [11, 13, 15, 16], as an emergent property of component interaction [12, 14, 17, 49], or as an automatically learned property from biased data [18, 19, 39–42, 88, 89, 91].”

“, using static or dynamic analyses for vacuity testing [8, 33] or to identify configuration options that cannot affect a test’s output [48], can directly improve efficiency of discrimination testing by identifying that changing a particular input characteristic cannot affect a particular test’s output, and thus no cau…”

“Already, software makes decisions in what products we are led to buy [53], who gets a loan [62], self-driving car actions that may lead to property damage or human injury [32], medical diagnosis and treatment [74], and every stage of the criminal justice system including arraignment and sentencing that determine who g…”

“, as a result of implementation bugs, as an unintended property of self-organizing systems [11, 13, 15, 16], as an emergent property of component interaction [12, 14, 17, 49], or as an automatically learned property from biased data [18, 19, 39–42, 88, 89, 91].”

“This dataset is well vetted: it has been used by others to devise discriminationfree algorithms [18, 88, 89], as well as for non-discrimination purposes [3, 43, 86].”

“Automated testing research has produced tools to generate tests, including random testing, such as Randoop [63, 64], NightHawk [4], JCrasher [20], CarFast [65], and T3 [69]; search-based testing, such as EvoSuite [29], TestFul [9], and eToc [78]; dynamic symbolic execution tools, such as DSC [37], Symbolic PathFinder…”

“Because developer-written test suites are often not as representative of field executions as developers would like [81], operational profiles can significantly improve the effectiveness of testing by more accurately representing real-world system use [7, 50] and the use of operational profiles has been shown to more a…”

“, as a result of implementation bugs, as an unintended property of self-organizing systems [11, 13, 15, 16], as an emergent property of component interaction [12, 14, 17, 49], or as an automatically learned property from biased data [18, 19, 39–42, 88, 89, 91].”

“history of warfare: For the first time, a collection of microchips and software will decide whether a human being lives or dies” [34].”

“Because developer-written test suites are often not as representative of field executions as developers would like [81], operational profiles can significantly improve the effectiveness of testing by more accurately representing real-world system use [7, 50] and the use of operational profiles has been shown to more a…”

“, using static or dynamic analyses for vacuity testing [8, 33] or to identify configuration options that cannot affect a test’s output [48], can directly improve efficiency of discrimination testing by identifying that changing a particular input characteristic cannot affect a particular test’s output, and thus no cau…”

“Then, statistical parity will deem the classifier fair, despite a discrepancy in how the applications are processed based on gender [13].”

“As the field matures, integrated toolkits are being developed, with the aim of making these methods more widely accessible and usable [7, 28, 35, 42, 70].”

“…in technical review articles (Berk et al. 2017, Chouldechova & Roth 2018, Friedler et al. 2018, Verma & Rubin 2018) and an in-progress textbook (Barocas et al. 2018), and has inspired a number of software packages (Galhotra et al. 2017, Zehlike et al. 2017, Angell et al. 2018, Wexler 2018).”