A Mutation-based Approach to Repair Deep Neural Network Models

Due to insufficient accuracy, defects in deep neural network models may cause serious consequences. How to detect the defects by testing and verification has attracted wide attention of researchers. However, there is a lack of effective automatic repair techniques for the defects. Most of existing deep neural network model repair methods are based on adversarial or coverage guidance to augment training data sets. Therefore, the cost of repair is relatively high, and different distributions are obviously shown between the augmented data and natural data, and the repair may cause overfitting. Inspired by the automatic repair techniques of traditional software, this paper proposes a mutation-based automatic repair approach for the deep neural network model to improve its accuracy. Our key idea is to rank the weights of the deep neural network model according to the influence on the test results, and then to adjust the weights in sequence based on the genetic algorithm to repair the defect.