Research And Application Of Adversarial Attack Integrated Defense Model Based On Service Orchestration

Artificial intelligence applications have achieved remarkable results in fields such as image recognition,autonomous driving,and finance.The application of cloud computing has accelerated the development speed of artificial intelligence applications and has become the primary way of artificial intelligence application deployment.However,existing research has shown that artificial intelligence applications based on deep neural networks are still facing serious threats from adversarial samples,and cloud computing also has data information security issues.As the application of deep neural networks becomes more and more widespread,the consequences will be disastrous.To defend against adversarial attacks,researchers have proposed an effective strategy,adversarial defense.But the current defense algorithms are mostly aimed at certain types of adversarial attacks,which are related to the experiments and empirical observations of specific neural network models and have poor generalization.Current research shows that there is still no single defense method which can have the best defense effect against all attacks.The integrated defense of statically bundling multiple defenders has been experimentally shown to fail to achieve better defense performance.This paper solves this problem from the architecture level and algorithm level,and the main contributions are summarized as follows:1.From the perspective of AI applications safety,combined with the AI engineering development process,we provide independent security services for AI applications.Research including:(1)An integrated defense architecture based on service orchestration security is proposed to provide independent security protection functions for artificial intelligence applications.(2)In actual business scenarios,provide data security protection for the artificial intelligence application of picture detection developed based on cloud services,and realize real-time,singlegrained detection of input data.2.On the one hand,Using the manifold hypothesis,learn the difference of manifold mapping between adversarial samples and legitimate samples.On the other hand,with reference to new ideas of service orchestration in microservices,service orchestration and the quantitative method of data statistics are used to solve the adversarial samples detection problem in image and network traffic and verify the effectiveness of the proposed method in experiments.the main works include the following points:(1)Using the manifold hypothesis,select the optimal image processing method for the adversarial example,arrange the service chain,and propose the image data integrated defense method based on image preprocessing method.(2)Combined with the preprocessing method of text data,select the optimal cutting length for the input sample,and propose the network traffic data integrated defense methods based on spatial filtering.