International Journal of Computer Science and Technology

International Journal of Computer Science and Technology | Vol. 3, No. 10, October 2022 | pp. 1–8

DOI: 10.46882/2022/IJCST/000213

Article Type: Original Research Paper

Title: Quantifying Adversarial Vulnerabilities in Graph Neural Networks via Targeted Topology Perturbation

Names of Authors: Wei-Dong Zhang¹, Amara Okafor²

Authors’ Affiliations: ¹School of Computer Science, Yangtze University of Technology, Wuhan, China; ²Department of Systems Engineering, University of Nigeria, Nsukka, Nigeria

Abstract: Graph Neural Networks (GNNs) are widely used to analyze relational structured applications like financial fraud detection maps and biological interactions. However, their structural dependency makes them highly vulnerable to malicious structural manipulations. This paper introduces an adversarial testing mechanism designed to evaluate GNN resilience by introducing targeted perturbations to node connections. The algorithm calculates node-centrality values alongside structural gradients to find and modify high-impact link structures with minimal modifications. We tested this attack method on standard benchmark network configurations, focusing primarily on GCN and GraphSAGE architectures. The assessment reveals that modifying a tiny fraction (less than 2.5%) of network links degrades node-classification accuracy by 38.4%. This drop demonstrates that minor topology edits can severely bypass standard neural classification security steps. To mitigate this vulnerability, we outline a robust structural defense strategy rooted in localized neighborhood-degree validation. This defense effectively neutralizes up to 72.0% of adversarial edge modifications during training operations.

Keywords: Graph Neural Networks, Adversarial Attacks, Topology Perturbation, Node Classification, Network Robustness, Machine Learning

Manuscript Timeline: Received: May 20, 2022; Revised: June 28, 2022; Accepted: July 22, 2022; Published: October 08, 2022

Subscribe to read the full article: https://internationalscholarsjournals.org/subscribe-to-read