Graph neural network model backdoor attack-oriented detection method and device

Abstract

The invention discloses a graph neural network model backdoor attack-oriented detection method and device, and the method comprises the steps: training a graph neural network model through employing graph data, so as to optimize the parameters of the graph neural network model; inputting the graph data into the parameter-optimized graph neural network model, calculating a loss function corresponding to the graph data, and performing reverse derivation on the loss function relative to an adjacent matrix of the graph data to obtain an importance degree value of each connecting edge to the loss function; extracting sub-graph structures with different connecting edge numbers according to the importance degree values, and dividing the sub-graph structures into a plurality of sub-graph libraries corresponding to the classification labels according to the classification labels; for each sub-graph library, calculating a distribution graph of the sub-graph structures according to the similarity between the sub-graph structures; and analyzing the similarity value in the distribution map corresponding to each sub-map library, and determining whether the map neural network model is attacked or not according to the similarity value. The backdoor attack detection of the graph neural network model is realized, and the security of the model is improved.

Description

technical field [0001] The invention belongs to the field of security detection, and in particular relates to a detection method and device for backdoor attack of a graph neural network model. Background technique [0002] In the process of graph neural network (GNN) solving graph evolution tasks, it also brings many problems, and the security issue of graph neural network model is a particularly important part in the whole process. Surprisingly, while there has been a lot of previous work on the security of DNNs for continuous data (e.g., images), little is known about the vulnerability of graph neural networks (GNNs) for discretely structured data (e.g., graphs), given its application With increased range, its safety is a high concern. In the process of completing downstream tasks, the graph neural network model needs a large amount of data sets to learn data set information, update model parameters, and make the model better complete downstream tasks. In this process, in...

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