IEMT: Inequality-Based Metamorphic Testing for Autonomous Driving Models

Abstract

In the last ten years, the development of deep learning has promoted the progress of autonomous driving. Several major manufacturers, including Google, Tesla, Baidu, Audi, etc., are building and actively testing self-driving cars. However, the safety of autonomous driving still raises concerns. Recent research has used metamorphic testing to evaluate the robustness of autonomous driving models, but metamorphic relations defined during the test are basically based on equality, and there are very few inequality-based metamorphic relations. Our goal is to provide more inequality-based metamorphic relations to efficiently detect autonomous driving model violations. IEMT proposes additional inequality-based metamorphic relations and compares the robustness of autonomous driving models based on different neural network models. The experimental results show that the metamorphic relations we proposed can detect inconsistent behaviors of the driving model quite efficiently.