Detection method, device, computer equipment and storage medium for network security

Abstract

The present invention relates to artificial intelligence, and provides a network security detection method, device, computer equipment and storage medium. The method includes obtaining a training set of a target network; determining an attack vector against a preset source network on the training set; determining a disturbance vector corresponding to the attack vector; constructing a generating network; training the generating network; The output vector of the device is used as the embedding vector, and the attack vector of the target network is calculated in the space where the embedding vector is located to obtain the third attack vector; the attack vector of the target network is input to the target network to detect the target network. In addition, the present invention also relates to blockchain technology, and the attack vector of the target network can be stored in the blockchain node. Through the present invention, the detection accuracy and search efficiency in the attack technology based on the black box can be improved.

Description

technical field [0001] The present invention relates to the technical fields of blockchain and artificial intelligence, in particular to a network security detection method, device, computer equipment and storage medium. Background technique [0002] The rapid development of deep neural network technology has made it widely used in various fields of artificial intelligence. However, existing research and practice have proved that deep neural networks are vulnerable to adversarial attacks. For the deep neural network model involved in computer vision tasks, small changes to image pixel values ​​that are not easy to be noticed will lead to wrong judgments in computer vision tasks such as image classification and target detection, which brings great challenges to computer vision tasks. big security risk. [0003] In order to reduce and eliminate these potential safety hazards, when defending against adversarial attacks, the attack image of the target application network is fi...

AI Technical Summary

Problems solved by technology

Black-box-based attack techniques in the prior art include migration-based technology and score-based technology. In this regard, the inventors found that migration-based technology uses a pre-trained source network with known model parameters to generate an attack Image, as the attack image of the target application network, but the effectiveness of generating the attack image in this way is low, so the accuracy of the reliability detection of the target application network is also low, and it is impossible to effectively detect the possible attacks on the target application network Vulnerabilities of the target application network; the score-based technique approximates the gradient of the target application network by sampling the input vector of the target application network and obtaining the output score value of the target application network, and then searches for attack images based on these gradient values. For this method, when the target When the dimension of the input vector of the application network is high, the output score value of the target application network needs to be queried multiple times, and the efficiency of searching for attack images is low, so the detection of the target application network cannot be efficiently completed

[0005] Therefore, when using the existing black-box-based attack technology to generate detection target application network security, there are problems of low detection accuracy or low search efficiency, which has become a technical problem that needs to be solved urgently in this field.

Images