To solve the path on the local explanation decision tree (Ribeiro et al., 2016) of a given seed sample to acquire a large number of diverse unfair samples. Symbolic Generation (Aggarwal et al., 2019) utilizes a constraint solver (Wang et al., 2018) AEQUITAS (Udeshi et al., 2018) integrates a global and local phase to search for unfair samples in the input space more systematically. For instance, THEMIS (Galhotra et al., 2017) first aims to measure the frequency of unfair samples by randomly sampling the value domain of each attribute. Testing problem of machine learning models. Multiple recent works (Galhotra et al., 2017 Udeshi et al., 2018 Aggarwal et al., 2019 Zhang et al., 2020) have investigated the fairness 1 1 1So far restricted to individual fairness. Most valuable test cases to mitigate the model's fairness issues. TheĮxperimental results confirm that our approach can effectively identify theįairness-related neurons, characterize the model's fairness, and select the Large-scale face recognition applications, i.e., VGGFace and FairFace. We have conducted experiments on widely adopted Important components enabling effective fairness testing of deep imageĬlassification applications: 1) a neuron selection strategy to identify theįairness-related neurons 2) a set of multi-granularity adequacy metrics toĮvaluate the model's fairness 3) a test selection algorithm for fixing theįairness issues efficiently. Propose DeepFAIT, a systematic fairness testing framework specifically designedįor deep image classification applications. Structured data or text without handling the high-dimensional and abstractĭomain sampling in the semantic level for image classification applications 2)įunctionality, i.e., they generate unfair samples without providing testingĬriterion to characterize the model's fairness adequacy. The following limitations: 1) applicability, i.e., they are only applicable for Existing fairness testing methods suffer from It is thus crucial to comprehensively test the fairness of theseĪpplications before deployment. Increasingly prevalent in our daily lives, their fairness issues raise more and Fairness Testing of Deep Image Classification with Adequacy MetricsĪs deep image classification applications, e.g., face recognition, become