Robustness Analysis Of Convolutional Neural Networks

With the emergence and development of machine learning,human beings hope to use computer belts instead of human beings to complete simple tasks with single task but huge workload,such as image classification.In this problem,convolutional neural network,especially deep convolutional neural network,is one of the most useful mainstream means at the present stage.Convolutional neural network has excellent performance advantages of high classification accuracy and high processing speed in the face of image problems.Meanwhile,with the continuous research and improvement of previous researchers,the existing convolutional neural network has been able to effectively process image classification problems.In this case,It is hoped to further discuss the robustness of convolutional neural networks in the face of interference.In this paper,the robustness problem is studied.Numerical experiments are carried out on two kinds of data sets.Four convolutional neural networks,three different kinds of noise and three ways of adding noise are used in the experiment.In addition to Alex Net,VGG Net and three-layer convolutional neural network,I also studied an improved model PCNN studied in my previous papers.In order to better observe the impact of each variation on the robustness of the network,the experiment was compared from three dimensions: the same network,the same noise analysis of the impact of different noise methods on the robustness;The influence of different noise on robustness was analyzed in the same network with the same noise addition mode.And the different robustness of different networks under the same noise and the same addition of noise.The numerical experimental results in this paper show that the image classification accuracy of several deep convolutional neural networks will not decrease by more than 10%regardless of the input of color images or gray images,and the deeper the structure of the network model is,the better the robustness of the network will be.However,the shallow neural network is limited by the depth of its own network structure and can only show strong robustness when the input is grayscale image.At the same time,the experimental results also show that the network has stronger robustness when it first learns disturbance and then faces noise during training.