Document Type

Conference Proceeding

Publication Date

12-2024

Publisher

Institute of Electrical and Electronics Engineers (IEEE)

Source Publication

GLOBECOM 2024 - 2024 IEEE Global Communications Conference

Source ISSN

2576-6813

Abstract

Fault detection is an integral part of the protection system in a power distribution network. Due to advanced computational capabilities, deep learning-based algorithms can significantly outperform traditional methods. However, these deep learning models are prone to adversarial attacks which are not well-addressed as traditional cyber attacks in distribution systems. More specifically, to capture the structure of distribution systems, graph neural networks (GNNs) are employed. Leveraging the backdoor attack model, we propose a novel graph-based adversarial attack algorithm for fault detection in power systems. It is further shown that the adaptable structure of GNN can make them vulnerable to adversarial attacks through graph motif insertion. The backdoor attack is carried out on IEEE-13 and IEEE-37 feeder systems and numerical results demonstrate a substantial decrease in fault detection accuracy as a result of these attacks.

Comments

Accepted version. GLOBECOM 2024 : IEEE Global Communications Conference, (2025): DOI. © 2024 Institute of Electrical and Electronics Engineers (IEEE). Used with permission.

Yahyasoltani_16708acc.docx (292 kB)
ADA Accessible Version

Share

COinS