GenMuNN: A mutation-based approach to repair deep neural network models

Abstract

Deep neural network (DNN) models have been widely used in e-commerce, games, automobiles, manufacturing, and so on. Improper structure, parameters, activation function, or incorrect loss function of the DNN models may cause defects in performance or security. As a result, there are some researches that focus on repairing DNN such as MODE and Apricot. However, the cost of repairing is high or the repair may lead to overfitting. In order to solve this problem, we propose GenMuNN, which is a Mutation-Based Approach to Repair Deep Neural Network Models. First, it analyzes the importance of the weights of the neurons in each layer of the DNN model to the correctness of the final prediction results, and ranks the weights according to the influence on the prediction results of the DNN model. Second, mutation is performed to generate mutants based on the rank of weights, and genetic algorithms are used to select mutants for the next round of mutation until the stop condition is touched. Experiments are carried on a set of DNN models which are trained with the MNIST dataset. The experimental results show that GenMuNN can improve the accuracy of the DNN models.