Fast Test Method of Intelligent Image Recognition System Based on Antagonistic Samples
Lilei Zhang, Jinyong Yao, Kai Wang · 2020
With the development of science and technology in the field of computer hardware and algorithm, deep neural network is widely used in the field of computer vision. However, intelligent image recognition systems have also encountered difficulties in reliability and safety. Especially when facing the counterexample, these systems cannot perform well. In this paper, the handwriting digital image recognition test is taken as an example with establishing a neural network of image recognition system. The difference of recognition performance between multi-layer network and single-layer network based on convolutional network is compared, and the confrontation sample and the test sample pair are used respectively. The network conducted tests to analyze the differences between them and concluded that tests based on countermeasures were more effective than tests based solely on test samples. And it is more efficient to find the reliability of intelligent image recognition system in recognition.