Research On Endogenous Control Technique Of Convolutional Neural Network

With the rapid development of cloud computing and artificial intelligence,there are more and more applications of AIaa S(Artificial Intelligence-as-a-Service)in the cloud.Among them,artificial intelligence application based on convolutional neural network is widely used.The application system based on convolutional neural network,like other information systems,which can be utilized by the enemy to bypass the existing access control mechanism of the system,has software and hardware vulnerabilities and security risks.The existing access control mechanism is difficult to effectively protect the convolutional neural network.Endogenous security is a security idea that integrates the security mechanism into the core processing parts of the protected system and is closely coupled with the information system.In response to the above problems,we integrat the idea of endogenous security into the core computing unit of convolutional neural network.The theory and method of endogenous control of convolutional neural network are first proposed,and an endogenous control model,two endogenous control algorithms and three optimization schemes are designed for the first time.The main research contents and innovations of this paper are as follows:1.The core computing structure of convolutional neural network is analyzed and summarized,which lays the foundation for the research of endogenous control technology of convolutional neural network.The endogenous control of convolutional neural network is the first fusion of the endogenous safety idea in the safety control of deep neural network model.Based on the core computing structure of convolutional neural network,we research and propose the key control points of the endogenous control model.Aiming at the problem that the existing access control mechanism can not effectively prevent the neural network model from losing control,we design and propose a convolutional neural network endogenous control model by closely coupling the endogenous control unit at the control point.2.Two algorithms are proposed about the convolutional neural network endogenous control model,which are single point endogenous control algorithm and mult-point endogenous control algorithm.The main body of the endogenous control unit is the endogenous control algorithm.By the close coupling of the endogenous control algorithm and the convolutional neural network operation,the endogenous control of the convolutional neural network is achieved.The single point endogenous control algorithm achieves the endogenous control at a single control point of the convolutional neural network and with fast speed and an average control rate of 68.88 percent.The multi-point endogenous control algorithm achieves the endogenous control at multiple control points of the convolutional neural network,which has better control effect but larger time cost.By changing the control strategy,this paper proposes a preliminary optimized multi-point endogenous control algorithm to reduce time overhead.3.In order to further improve the safety control effect of the endogenous control algorithm on the convolutional neural network,three optimization schemes are proposed.The average control rate of the endogenous control algorithm applied to the weight control points of convolutional neural network is almost 0%.Therefore,this paper proposes an optimization scheme of the endogenous control algorithm based on the mean collapse,which is applied to the weight control points and make the average control rate reach 100%.When the endogenous control algorithm is applied to the bias control points of convolutional neural network,the average control rate of the bias control points in the middle layer is lower than that in the deeper layer.Therefore,this paper proposes an optimization scheme based on the giant parameter,which is applied to the bias control point,and can effectively improve the average control rate of the endogenous control algorithm applied to the bias control point in the middle operation layer.Finally,we design and propose an optimization scheme based on reverse activation.This optimization scheme,which is applied in the activation function control point,can improve the average control rate of the endogenous control algorithm.Three optimization schemes are suitable for two endogenous control algorithms.The convolution network endogenous control algorithm based on the optimization scheme can achieve the complete control of the convolution neural network.