Research And Application Of A Generative Adversarial Algorithm And Its Impact On Adversarial Examples

Recent research shows that some machine learning algorithms still face many security threats,which affect the system security of practical applications based on deep learning.By researching two important concepts in adversarial deep learning: adversarial attacks and generative adversarial networks,and with the combination of generative adversarial networks and adversarial attacks to explore the adversarial training application scheme of deep learning algorithm.This paper completes the research of one generative adversarial algorithm and its impact on adversarial examples.The whole paper can be divided into three sections which generally base on the core concept of adversary.In the first section,we explored a conditional generative adversarial capsule networks,and successfully transferred the excellent feature learning capabilities of the capsule network to the training process of the generative adversarial network.The experiment proved its not only stably produce high-quality images in the image generation task,but also inhibit the occurrence of pattern collapse more effectively,then provides beneficial insights for the design of superior learning model.In the second part,based on the design of generative adversarial network and relevant insights,and with the help of generative adversarial network's excellent learning ability on data distribution,a generic adversarial examples generation algorithm that is not limited by original data is proposed.The experiment proved its not only generates high-quality attack examples breaking the limits of the original data,but also has good aggression and attack migration competence,then provides more possibilities for adversarial training and improving the robustness of deep neural network.In the third part,considering the shortcomings of the traditional deep learning training system,we apply the generic adversarial example generation algorithm designed in the second part to the training system,and put forward a more stable deep learning training system based on generative adversarial network,which can obtain a more defensive network model and help the safe application of neural network.This paper explores the adversarial concept from three aspects: the principle of example generation,the algorithm design of the adversarial example and the application of the adversarial example algorithm,and conducts a lot of experiments and analysis on MNIST and CIFAR-10 Datasets,which provides a solution for further establishing a safe deep learning model.