Targeted Label Adversarial Attack on Graph Embedding

Graph embedding is a popular technique used in various real-world applications to learn low-dimensional representations for nodes or edges in a graph. The increasing interest in graph mining has led to the development of attack methods on graph embedding. Most of these attack methods aim to generate perturbations that maximize the deviation of prediction confidence. However, they often struggle to accurately misclassify instances into the desired target label, and the non-minimized perturbations can be easily detected by defense methods. To address these challenges, we propose a novel targeted label adversarial attack on graph embedding. It leverages classification boundary and gradient information in the target graph embedding method to generate adversarial graphs for attacking graph embedding methods. The approach first estimates the classification boundaries for different categories and then calculates the minimum perturbation matrix required to misclassify the targeted node based on the target classification boundary. Finally, the adjacency matrix is modified based on the maximum absolute value of the perturbation matrix. Extensive experiments demonstrate that our method achieves state-of-the-art attack performance with minimal perturbations.