Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RW-NSGCN: A Robust Approach to Structural Attacks via Negative Sampling

Shuqi He, Jun Zhuang, Ding Wang, Jun Song

TL;DR

RW-NSGCN addresses the vulnerability of graph neural networks to topology perturbations and weight instability by fusing Random Walk with Restart and PageRank-based negative sampling with a Determinantal Point Process–driven GCN. The approach constructs non-neighbor candidate sets across multiple path lengths, scores them with global and local signals, and enforces diversity via DPP before integrating them into a GCN as negative samples. Empirical results on Cora, CiteSeer, and PubMed show superior accuracy and robustness to structural attacks, with ablations confirming the benefits of combining RWR, PageRank, and DPP. The work offers a practical, robust framework for secure graph learning in dynamic networks, with clear guidance on parameter choices and attack resilience.

Abstract

Node classification using Graph Neural Networks (GNNs) has been widely applied in various practical scenarios, such as predicting user interests and detecting communities in social networks. However, recent studies have shown that graph-structured networks often contain potential noise and attacks, in the form of topological perturbations and weight disturbances, which can lead to decreased classification performance in GNNs. To improve the robustness of the model, we propose a novel method: Random Walk Negative Sampling Graph Convolutional Network (RW-NSGCN). Specifically, RW-NSGCN integrates the Random Walk with Restart (RWR) and PageRank (PGR) algorithms for negative sampling and employs a Determinantal Point Process (DPP)-based GCN for convolution operations. RWR leverages both global and local information to manage noise and local variations, while PGR assesses node importance to stabilize the topological structure. The DPP-based GCN ensures diversity among negative samples and aggregates their features to produce robust node embeddings, thereby improving classification performance. Experimental results demonstrate that the RW-NSGCN model effectively addresses network topology attacks and weight instability, increasing the accuracy of anomaly detection and overall stability. In terms of classification accuracy, RW-NSGCN significantly outperforms existing methods, showing greater resilience across various scenarios and effectively mitigating the impact of such vulnerabilities.

RW-NSGCN: A Robust Approach to Structural Attacks via Negative Sampling

TL;DR

RW-NSGCN addresses the vulnerability of graph neural networks to topology perturbations and weight instability by fusing Random Walk with Restart and PageRank-based negative sampling with a Determinantal Point Process–driven GCN. The approach constructs non-neighbor candidate sets across multiple path lengths, scores them with global and local signals, and enforces diversity via DPP before integrating them into a GCN as negative samples. Empirical results on Cora, CiteSeer, and PubMed show superior accuracy and robustness to structural attacks, with ablations confirming the benefits of combining RWR, PageRank, and DPP. The work offers a practical, robust framework for secure graph learning in dynamic networks, with clear guidance on parameter choices and attack resilience.

Abstract

Node classification using Graph Neural Networks (GNNs) has been widely applied in various practical scenarios, such as predicting user interests and detecting communities in social networks. However, recent studies have shown that graph-structured networks often contain potential noise and attacks, in the form of topological perturbations and weight disturbances, which can lead to decreased classification performance in GNNs. To improve the robustness of the model, we propose a novel method: Random Walk Negative Sampling Graph Convolutional Network (RW-NSGCN). Specifically, RW-NSGCN integrates the Random Walk with Restart (RWR) and PageRank (PGR) algorithms for negative sampling and employs a Determinantal Point Process (DPP)-based GCN for convolution operations. RWR leverages both global and local information to manage noise and local variations, while PGR assesses node importance to stabilize the topological structure. The DPP-based GCN ensures diversity among negative samples and aggregates their features to produce robust node embeddings, thereby improving classification performance. Experimental results demonstrate that the RW-NSGCN model effectively addresses network topology attacks and weight instability, increasing the accuracy of anomaly detection and overall stability. In terms of classification accuracy, RW-NSGCN significantly outperforms existing methods, showing greater resilience across various scenarios and effectively mitigating the impact of such vulnerabilities.
Paper Structure (21 sections, 15 equations, 2 figures, 3 tables, 1 algorithm)

This paper contains 21 sections, 15 equations, 2 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Preliminaries
  4. menthod
  5. Definition
  6. Overview
  7. Selection of Negative Samples
  8. GCN Based on Label Propagation with DPP Sampling
  9. experiment
  10. Datasets
  11. Baselines
  12. Experiment Settings
  13. Metrics
  14. Accuracy
  15. Mean Average Distance
  16. ...and 6 more sections

Figures (2)

  • Figure 1: The bar chart illustrates the accuracy of three representative models in classifying network structures after two types of artificial perturbations: topological perturbation (CTBCA) and weight perturbation (TWPA). 'S' denotes the SDGCN model, the state-of-the-art model, 'R' represents the RW-NSGCN model developed in this study, and 'G' stands for the classic GCN model. The horizontal axis represents classification models, and the vertical axis represents accuracy. The height of the bars indicates the accuracy of each model under specific conditions.
  • Figure 2: This figure illustrates the comparative accuracy of node selection using different $L$ values ($L=5$ and $L=6$) on the Cora and Citeseer datasets. Here, $L$ represents the maximum distance moved to select non-neighboring nodes. The left image shows that in the Cora dataset, the model with $L=5$ demonstrates significantly higher accuracy, despite some fluctuations. Meanwhile, the right image indicates that in the Citeseer dataset, the model with $L=5$ displays slightly higher accuracy than $L=6$ but with greater variability.