Anti-immune method for improving robustness on graph data

Abstract

The invention provides a method for protecting graph data, wherein the graph data comprises nodes and edges between the nodes, and the nodes have classification labels and attributes. The method comprises the following steps: step 100: selecting key node pairs from graph data; and step 200, carrying out immune protection on the selected key node pairs, wherein the immune protection comprises the steps of protecting connecting edges between the node pairs with high attribute similarity in the same category and keeping no connecting edges between the node pairs with low attribute similarity in different categories. According to the invention, the data-oriented anti-immune method is put forward for the first time, a training process of a GNN model does not need to be influenced, any optimization process is not needed, and a large amount of calculation power and time needed by optimization are saved; and anti-immunization directly vaccinates the key node pairs to protect the key node pairs from being attacked and damaged, so that robustness of the whole graph is improved, and the performance of the GNN on a clean graph is ensured.

Description

technical field [0001] The invention relates to the field of graph data mining, and more particularly relates to the protection of graph data. Background technique [0002] Graph data is ubiquitous in real life and is used to represent entities and complex relationships between entities. Typical graph data include: social networks, citation networks, transportation networks, biological networks, etc. The graph can be represented as such a structure: graph G=V,E) contains two sets: node set V and edge set E, where each node t∈V has its own attribute a t and label y t , the above structure is graph data. Taking a social network blog as an example, each node represents a post, the attribute of the node is the text word vector in the post, the label of the node is the community where the post is located, and each edge represents the relationship between posts. [0003] In recent years, a number of powerful graph data modeling tools have emerged, namely Graph Neural Networks (...

AI Technical Summary

Problems solved by technology

Of the above methods, the first two can achieve the effect of detecting attacks or authenticating robustness, but it is difficult to improve the robustness. The last one can effectively defend against attacks, but this method can only take effect after retraining the GNN model. Still need a lot of optimization calculations

Images