Mutation Operator Reduction for Deep Learning System

The mutation testing method of Deep Learning (DL) system proposes a series of DL mutation operators, but the same as traditional software mutation testing methods, a large number of mutants will be generated during the testing process, which will cause huge costs. The traditional mutation operator reduction method is based on source program business logic. Owe to the fundamental difference between traditional system and DL system, traditional reduction methods cannot be directly applied to the DL mutation operators. In this paper, we propose the mutation operator reduction method for DL system, which can be divided into three steps. It firstly classifies all mutation operators by the scope of action of them. Then, it combines different classes of mutation operators. Finally, it analyzes the mutation score of different mutation operators combinations to obtain a sufficient mutation operators subset. This method has been tested on the MNIST datasets and the LENET-5 model. The experimental results shows that the number of mutants reduced by 41.67%, which effectively proved that our reduction method can effectively reduce the number of mutants generated, reduce the testing cost, and improve the accuracy of the mutation score.