深度神经网络模型鲁棒性测试方法及应用OA

To address the lack of effective robustness testing methods for Deep Neural Network(DNN)models of intelligent systems,this paper first introduces the definition of DNN model robustness,and proposes robustness test evaluation indicators such as disturbance stability and performance fluctuation degree.Then,robustness testing methods for DNN models are proposed from three aspects:noise interference,data distribution and extreme data.Finally,a case application is carried out on the You Only Look Once version 5(YOLOv5)algorithm model for target detection of a certain ground unmanned platform,which verifies the effectiveness and feasibility of the proposed methods.