A robustness testing method for neural networks based on adversarial example generation techniques

Neural network robustness testing is crucial for ensuring the safety and stability of intelligent systems, with identifying the robustness boundary of models being one of its core tasks. Adversarial example generation has emerged as a novel technical approach to approximate this boundary. This paper begins by providing formal definitions of neural network robustness and adversarial examples, summarizing seven common types of adversarial example attack scenarios, and elaborating on five adversarial example generation methods, including L-BFGS, FGSM, and BIM. Subsequently, robustness boundary testing is conducted on a 4-layer fully connected neural network (DNN) and the convolutional neural network LeNet-5 using the MNIST handwritten digit image dataset and the widely applied FGSM method. The results indicate that, with a model failure threshold set at an accuracy rate below 80%, the approximate robustness boundary <i>&epsilon;</i> = 0.005 for DNN shows a drop in prediction accuracy from 97.48% to 81.62%, while <i>&epsilon;</i> = 0.125 for LeNet-5, the accuracy decreases from 97.48% to 80.47%, demonstrating that LeNet-5's robustness is significantly superior to that of DNN. The paper concludes with a brief overview of methods to defend against adversarial example attacks, offering references for enhancing neural network robustness based on adversarial example generation techniques.