Home // DATA ANALYTICS 2025, The Fourteenth International Conference on Data Analytics // View article

Trigger Injection via Clustering for Backdoor Attacks on Heterogeneous Graphs

Authors:
Honglin Gao
Lan Zhao
Gaoxi Xiao

Keywords: heterogeneous graph; backdoor attack.

Abstract:
Heterogeneous graph neural networks have achieved remarkable success in modeling multi-relational data. However,the risks associated with backdoor attack have largely gone unexplored. In this paper, we present a new structure-based backdoor attack method for heterogeneous graph neural networks. Our method uses a set of designed trigger nodes in the graph connected to semantically related parts of the graph using clustering-based trigger node selection. Triggering nodes cause the model to misclassify certain target nodes as an attacker specified class while still keeping a high accuracy on the clean data. Preliminary experiments on publicly available benchmark datasets show that our proposed backdoor attack is effective and stealthy. This shows that there is a clear need for security awareness in heterogeneous graph learning.

Pages: 15 to 17

Copyright: Copyright (c) IARIA, 2025

Publication date: September 28, 2025

Published in: conference

ISSN: 2308-4464

ISBN: 978-1-68558-293-7

Location: Lisbon, Portugal

Dates: from September 28, 2025 to October 2, 2025