Differential privacy aggregation-based graph neural network construction method and construction system

Abstract

The invention relates to a differential privacy aggregation-based graph neural network construction method and system, and the method comprises the steps: firstly obtaining a graph data set G from a crowdsourcing platform, and initializing a graph neural network model which at least comprises two layers of network structures; inputting the graph data set G into the graph neural network model, performing differential privacy aggregation processing on each node data of the graph data set G in the first-layer network structure, and outputting the data to the second-layer network structure; taking the output of the first-layer network structure as the input of a second-layer network structure, performing aggregation and classification prediction, and calculating a loss function by using a classification prediction result and a true value; and according to the value of the loss function, carrying out iterative training and model parameter updating or ending training, and outputting a current model parameter. In the model construction process, disturbance information is added to aggregated information by using the idea of differential privacy aggregation so as to achieveprotection of user privacy information in a data set.

Description

technical field [0001] The invention relates to the technical field of network security, in particular to a method and system for constructing a graph neural network based on differential privacy aggregation. Background technique [0002] Graph neural networks (GNNs) have become a standard tool for analyzing and learning from graph data and have been successfully applied in various fields. However, high-performance GNNs models often contain rich node features and edge information, which may leak some sensitive information of the dataset during model training. [0003] During model training, large-scale graph-structured data are typically collected from crowdsourcing platforms, which often encode sensitive information not only about individuals but also interactions with them, making direct release rather insecure. For example, in a social network like Facebook, interest groups should be private with varying levels of privacy in a user's friends list (eg or comments). Secon...

AI Technical Summary

Problems solved by technology

Therefore, when the graph neural network model constructed based on these private information is targetedly attacked, it is very easy to cause the leakage of sensitive private data of network users.

Images