A Gradient-Based Graph Adversarial Example Generation Method by Adding False Nodes for Document Classification

Abstract

The invention discloses a gradient-based method for generating graph adversarial samples by adding false nodes, including: (1) obtaining original graph data and graph node classification models to construct an attack target set; adding false nodes to the original graph data to obtain adversarial sample; (2) select the attack target from the attack target set (v * ,y * ); (3) Input the current adversarial sample into the classification model, and use the loss function to calculate the loss; (4) Calculate the gradient of the loss to the input adjacency matrix, and select the node corresponding to the element with the largest gradient value in the row corresponding to the false node Connect with the added false node to get a new adversarial sample; (5) input the new adversarial sample into the classification model, if the classification result is y * , the new adversarial example is the generated adversarial example, otherwise skip to step (3). Utilizing the present invention, the generated adversarial examples can effectively affect the classification results of the graph deep learning model, and are more feasible in practical applications.

Description

Technical field [0001] The present invention belongs to the field of artificial intelligence information security technology, in particular, relating to a gradient based on the graph of the graph to which the fake node is added to the anti-sample generation method. Background technique [0002] The figure in the chart is a graphic constructed from a number of given points and two points of connection, which is often used to describe some specific relationship between certain things, representing things, two points The line indicates a certain relationship between the two things. The figure G is an ordered binary group (V, E), where V is called a vertex set, that is, a set of all vertices in the figure, e calls the edge set, that is, the edges between all vertices set. Simply put, the vertex indicates something, and the relationship between things is indicated. In addition, the attribute graph may be represented as an ordered binary group (A, X), where A is referred to as an adjac...

AI Technical Summary

Problems solved by technology

In the field of graph data, current research is to make the classification results of target nodes mislead by adding or deleting existing edge or node features.

But this method may be difficult to implement in actual scenarios. For example, in a social network, if you want to delete or add an edge between two users, you may need to obtain the login permissions of these users, but this is difficult to obtain in actual situations.

Images