Detection method and device for graph neural network model backdoor attack

Abstract

The invention discloses a detection method and device for a graph neural network model backdoor attack, comprising: using graph data to train the graph neural network model to optimize the parameters of the graph neural network model; inputting the graph data into the parameter-optimized graph In the neural network model, the loss function corresponding to the graph data is calculated, and the reverse derivative of the loss function relative to the adjacency matrix of the graph data is obtained to obtain the importance value of each edge to the loss function; different edge values ​​are extracted according to the importance value According to the classification label, the sub-graph structure is divided into multiple sub-graph galleries corresponding to the classification labels; for each sub-graph gallery, the distribution map of the sub-graph structure is calculated according to the similarity between the sub-graph structures; analysis of each The similarity value in the distribution graph corresponding to each sub-gallery, and determine whether the graph neural network model is attacked according to the size of the similarity value. Realize backdoor attack detection on the graph neural network model to improve the security of the model.

Description

Technical field [0001] The present invention belongs to the field of security detection, specifically relates to a method and apparatus for detecting a backdoor attack oriented to a graph neural network model. Background [0002] Graph neural networks (GNNs) also pose many problems in solving graph evolution tasks, and the security issues regarding graph neural network models are a particularly important part of the overall process. Surprisingly, despite a lot of previous security work on DNNs for continuous data (e.g., images), little is known about the vulnerability of graph neural networks (GNNs) for discrete structural data (e.g., graphs), and its security is a highly worrisome issue given its increased range of applications. In the process of completing the downstream task, the graph neural network model needs a large number of data sets to learn the data set information, update the model parameters, and make the model better complete the downstream tasks, and the initial da...

AI Technical Summary

Problems solved by technology

[0004] In the field of graph neural network backdoor attack, different attack methods have emerged, but the detection method for backdoor attack is still in a blank state, which makes the graph neural network model always in danger of suffering from backdoor attack in terms of security. , causing serious consequences

Images