Automatic Fairness Testing of Neural Classifiers Through Adversarial Sampling

Abstract

Although deep learning has demonstrated astonishing performance in many applications, there are still concerns about its dependability. One desirable property of deep learning applications with societal impact is fairness (i.e., non-discrimination). Unfortunately, discrimination might be intrinsically embedded into the models due to the discrimination in the training data. As a countermeasure, fairness testing systemically identifies discriminatory samples, which can be used to retrain the model and improve the model's fairness. Existing fairness testing approaches however have two major limitations. Firstly, they only work well on traditional machine learning models and have poor performance (e.g., effectiveness and efficiency) on deep learning models. Secondly, they only work on simple structured (e.g., tabular) data and are not applicable for domains such as text. In this work, we bridge the gap by proposing a scalable and effective approach for systematically searching for discriminatory samples while extending existing fairness testing approaches to address a more challenging domain, i.e., text classification. Compared with state-of-the-art methods, our approach only employs lightweight procedures like gradient computation and clustering, which is significantly more scalable and effective. Experimental results show that on average, our approach explores the search space much more effectively (9.62 and 2.38 times more than the state-of-the-art methods respectively on tabular and text datasets) and generates much more discriminatory samples (24.95 and 2.68 times) within a same reasonable time. Moreover, the retrained models reduce discrimination by 57.2% and 60.2% respectively on average.

Fairness Testing of Deep Image Classification with Adequacy Metrics

“…The first question is how to realize the domain transformation function 𝑇 𝐴→𝐵 . Note that this is straightforward for structured or text data which can be done by replacing the protected feature or token with a value from a predefined domain [55,56]. However, for image data, the sensitive attribute of interest is hidden from the input feature space.…”

“…The above metrics enables us to evaluate a model's fairness adequacy. The follow-up questions are 1) how to generate diverse test cases to improve the fairness adequacy, and then 2) how to select the most valuable test cases to enhance the model's fairness by augmented training, which has been proved to be useful in enhancing the robustness or fairness of DNN [25,55,56].…”

“…Gradient-based Generation (GG) It is noticeable that gradient is an effective tool to generate test cases for DL model [16,29,55,56].…”

“…We utilize the gradient-based method in [55,56] to approximate the solution of the optimization problem 16, which relies on the sign of gradient of loss function with respect to the input, i.e., 𝑠𝑔 = 𝑠𝑖𝑔𝑛(▽ 𝑥 𝐽 (𝑥, 𝑦)) and 𝑠𝑔 ′ = 𝑠𝑖𝑔𝑛(▽ ′ 𝑥 𝐽 (𝑥 ′ , 𝑦)). As shown in Equation 18, we then choose the pixel with the same direction (sign) and the corresponding direction for perturbation.…”

“…Symbolic Generation [1] utilizes a constraint solver [49] to solve the path on the local explanation decision tree [36] of a given seed sample to acquire a large number of diverse unfair samples. State-of-the-art work ADF [55,56] and its variants [3] adopt a gradient-guided search strategy to identify unfair samples more effectively. Despite the considerable progress, existing fairness works still suffer from the following limitations: 1) applicability, i.e., they are only applicable for structured data or text without handling the highdimensional and abstract domain sampling in the semantic level for image classification applications; 2) functionality, i.e., they generate unfair samples without providing testing criterion to characterize the model's fairness adequacy.…”