Testing Artificial Intelligence System Towards Safety and Robustness: State of the Art

With the increasing development of machine learning, conventional embedded systems cannot meet the requirement of current academic researches and industrial applications. Artificial Intelligence System (AIS) based on machine learning has been widely used in various safety-critical systems, such as machine vision, autonomous vehicles, collision avoidance system. Different from conventional embedded systems, AIS generates and updates control strategies through learning algorithms which make the control behaviors nondeterministic and bring about the test oracle problem in AIS testing procedure. There have been various testing approaches for AIS to guarantee the safety and robustness. However, few researches explain how to conduct AIS testing with a complete workflow systematically. This paper provides a comprehensive survey of existing testing techniques to detect the erroneous behaviors of AIS, and sums up the involved key steps and testing components in terms of test coverage criterion, test data generation, testing approach and common dataset. This literature review aims at organizing a standardized workflow and leading to a practicable insight and research trend towards AIS testing.