Proj CDeepFuzz Paper Reading: Astraea: Grammar-based fairness testing

Abstract

Background:

1. discriminatory inputs, e.g., from societal bias, produce error, need to conduct fairness testing(generating discriminatory inputs that reveal and explain biases)

本文：Astraea
Github: https://github.com/sakshiudeshi/Astraea
Task: leverages context-free grammars to generate discriminatory inputs that reveal fairness violations in software systems
方法：

1. probabilistic context-free grammars
2. provides fault diagnosis by isolating the cause of observed software bias

实验：
Dataset: 18 software systems that provide three major natural language processing (NLP) services-Coreference Resolution, MLM, Sentiment Analysis
效果：

1. generated fairness violations at a rate of about 18%
2. ASTRAEA generated over 573K discriminatory test cases and found over 102K fairness violations
3. improves software fairness by about 76% via model-retraining, on average