DeepGemini: Verifying Dependency Fairness for Deep Neural Network

Deep neural networks (DNNs) have been widely adopted in many decision-making industrial applications. Their fairness issues, i.e., whether there exist unintended biases the DNN, receive much attention and become critical concerns, which can directly cause negative impacts our daily life potentially undermine of society, especially with their increasing deployment at an unprecedented speed. Recently, some early attempts made to provide assurance DNNs, such as testing, aims finding discriminatory samples empirically, certification, develops sound but not complete analysis certify DNNs. Nevertheless, how formally compute scores (i.e., percentage fair input space), is still largely uninvestigated. In this paper, we propose DeepGemini, a novel formal technique for contains two key components: sample discovery score computation. To uncover samples, encode DNNs safety properties search by means state-of-the-art verification techniques This reduction enables us be first samples. score, develop counterexample guided analysis, utilizes four heuristics efficiently approximate lower bound score. Extensive experimental evaluations demonstrate effectiveness efficiency DeepGemini on commonly-used benchmarks, outperforms DNN certification approaches terms both scalability.