Fuzzing for Deep Learning Models

Deep learning models have been widely used in security fields such as autonomous vehicles, and the testing of their quality problems has gradually attracted attention. Fuzzing has become an important testing method because of its efficient fault revealing ability. The quality and effectiveness of test cases generated by existing fuzzing methods are not high. In this paper, we propose a fuzzing method for deep learning models that uses heuristic policy mutation inputs to generate test cases. Thus, the quality of test cases is improved, so that the classification error of the model can be tested faster. By testing the image classification model, the experiments show that the generated test cases improve the neuron coverage of the model, and the time to find the model classification error is shorter.