Mutation Testing for Artificial Neural Networks: An Empirical Evaluation

摘要

Testing AI-based systems and especially when they rely on machine learning is considered a challenging task. In this paper, we contribute to this challenge considering testing neural networks utilizing mutation testing. A former paper focused on applying mutation testing to the configuration of neural networks leading to the conclusion that mutation testing can be effectively used. In this paper, we discuss a substantially extended empirical evaluation where we considered different test data and the source code of neural network implementations. In particular, we discuss whether a mutated neural network can be distinguished from the original one after learning, only considering a test evaluation. Unfortunately, this is rarely the case leading to a low mutation score. As a consequence, we see that the testing method, which works well at the configuration level of a neural network, is not sufficient to test neural network libraries requiring substantially more testing effort for assuring quality.