A Coverage-Guided Fuzzing Framework based on Genetic Algorithm for Neural Networks

Abstract

Due to the inherent difference between neural network and traditional software, it is very difficult to test it. At present, the use of fuzzing methods may be an effective exploration direction. We choose coverage-guided fuzzing as a method to test neural networks, and use neuron coverage as a coverage metric during execution. The effectiveness of neuron coverage will be demonstrated through experiments. On this basis, we designed a genetic algorithm-based fuzzing framework for neural networks, attempting to achieve greater coverage in a shorter time. And through the method of experimental comparison, the test efficiency of the framework is verified.