Bridging Fuzz Testing and Metamorphic Testing for Classification of Machine Learning

Abstract

Artificial Intelligence (AI) built-in Consumer Electronics is popular, but it is hard to test and evaluate AI-based system with the existing performance metrics. Even though AI-based systems are implemented in software with flexibility, bias and non-determinism property etc., they can suffer the same defects as other software. That is why new software testing approaches are needed when testing AI-based systems. Therefore, this paper proposes a bridging approach between fuzz testing and metamorphic testing focus on the classification of machine learning. This approach can be used as a test oracle for classification of training data.