Research On Class Activation Mapping Visualization And Adversarial Example Generation Algorithm For CNN

In recent years,deep learning technology has achieved great success in the field of computer vision.However,the model based on deep learning technology has security problems that can be attacked by adversarial examples.Attacker adds a specific perturbation to the original example which can make the model output error results with high confidence.Studying adversarial examples can inspire people to think about the security risks of deep learning technology and eliminate them.Existing adversarial examples generation methods usually focus on classification loss and regression loss of the target model.Adversarial examples are highly coupled to the model,which makes the adversarial examples less transferable between different models.From the perspective of interpretability of convolutional neural networks,a CAM-based adversarial examples generation method is proposed to improve the transferability of adversarial examples.The main work of this dissertation is as follows:1.A novel method called SL-CAM(Score-weighted and Layer-wise Class Activation Mapping)is proposed.SL-CAM generates class activation maps with high granularity by fusing the convolutional feature maps from shallow to deep layers of the network model.SL-CAM breaks the conventional idea that the current class activation mapping algorithm only uses the convolutional feature map of the highest layer.Firstly,gradient and activation information is used to generate class activation map of each layer.Secondly,the class activation map of each layer is integrated into the input example as a position mask to obtain the change of current category confidence.Finally,based on the change of the confidence,the class activation map of each layer is weighted to obtain a class activation map that fuses the features of all layers.Experimental results show that SL-CAM outperforms current class activation mapping algorithms in terms of average increase rate,average drop rate,consistency index,and image fine-grainedness.2.An adversarial example generation algorithm based on class activation map is proposed.In the image classification task,we realize the generation of adversarial examples under L∞ constraint.The algorithm mainly includes two parts:loss function design and iteration optimization strategy design.Firstly,in addition to the conventional category loss,attention loss and coupling degree loss with the original example are added into the loss function.Secondly,momentum iteration method is adopted and the noise perturbation is partially erased with a certain probability in each iteration update.Adversarial attack experiments are conducted on MNIST,Cifar10,and ISLVRC 2012 with different models.The experiments show that the generated adversarial examples can perform better in terms of white-box and transferability.In the object detection task,we realize the generation of adversarial examples under L0 constraint.Firstly,it obtains the location region of L0 constraint by CAM.Secondly,the confidence loss,attention loss and coupling loss are used in the loss function.Finally,the iteration optimization strategy is consistent with that proposed in the image classification task.Experiments are conducted on MS-COCO2017 dataset.The attack effect of object disappearing is successfully achieved and generated adversarial examples have better transferability.