Research And Design Of Computing Resource Management System In Cloud Service Platform

Nowadays,Artificial intelligence advancing fast,but all kinds of hacking are waiting to be rampant.Machine learning algorithms are becoming increasingly powerful,but at the same time,information privacy leakage,the spread of malicious code,network attacks and other behaviors are also rampant.Although deep learning has been applied in some fields,such as security vulnerability detection,Web application firewall and virus detection,hacker attack means have also changed accordingly.The relationship between deep learning and hacking is a chess-game as they evolve from one another.Face recognition,face payment has been popular in daily life.In order to improve the security of deep learning in the field of image security,it is necessary to understand the process of image classification,to deeply investigate the principle of anti-attack methods,and comprehensively improve the defensive performance of deep learning models.Aiming at all these problems,this paper studies the single target image classification model,multi-target target detection model,anti-attack method and antidefense system in deep learning.The main work of this paper is as follows:1.Defogging processing was introduced to make the dataset of single target traffic signs,and an image classifier based on convolutional neural network was built to realize the recognition of single target traffic signs,with an accuracy rate of 87%.The multi-objective traffic sign dataset was made by introduced defogging processing,and the YOLOV5-based target detection method was built and optimized to achieve the multi-objective traffic sign recognition,with an accuracy rate of 95.65% after the training was completed;2.Analyzed the against attack toward recognition according to single target traffic signs classifier,realised the optimized C&W algorithm of the white box attack,and the Fast Gradient Sign Method(FGSM)and depth cheating algorithm(Deep Fool)based on gradient optimization,realized the attack on image classifier to achieve the effect of making image classifier misjudgment;The against attack of the multi-objective traffic sign recognition target detector recognition are studied,realized the darkness attack,brightness attack,gaussian noise attack,salt and pepper noise attack,FGSM attack and adversarial patch attack,to achieve to attack detector and achieve the effect of the framework failure or classification misjudgment;3.The robustness in image transformation of the sample from single target traffic sign recognition against attack image classifier and the original sample is analyzed,and the influence results of image rotation,filter,contrast and brightness,Gaussian noise,pepper and salt noise changes on the model are obtained.The performance of Defense based on DefenseGAN network is analyzed,and the result of its influence on the model is obtained.Build against attack,brightness and dark patch attack,gaussian noise attack,salt and pepper noise,fast gradient descent attacked in multi-objective traffic sign recognition combat training,compare the influence of the robust performance of different attack methods to model,and draw conclusion all can enhance the robust performance target detector after training.