Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robustness questions the interpretability of graph neural networks: what to do?

Kirill Lukyanov, Georgii Sazonov, Serafim Boyarsky, Ilya Makarov

TL;DR

This work addresses the intertwined challenges of robustness and interpretability in graph neural networks under adversarial threats. It introduces a comprehensive, open-source benchmark that evaluates six GNN architectures across five datasets using four interpretability metrics (Fidelity, Stability, Consistency, Sparsity) while applying several defenses against poisoning and evasion attacks. The study finds that Fidelity and Consistency are relatively stable across settings, whereas Sparsity and Stability are more sensitive to domain, architecture, and defense choices; most defenses tend to improve interpretability, though effects vary by method and model. By revealing nuanced interactions between robustness and interpretability, the paper provides a practical foundation for designing GNNs that are both trustworthy and resilient, and it offers a reusable framework for future research.

Abstract

Graph Neural Networks (GNNs) have become a cornerstone in graph-based data analysis, with applications in diverse domains such as bioinformatics, social networks, and recommendation systems. However, the interplay between model interpretability and robustness remains poorly understood, especially under adversarial scenarios like poisoning and evasion attacks. This paper presents a comprehensive benchmark to systematically analyze the impact of various factors on the interpretability of GNNs, including the influence of robustness-enhancing defense mechanisms. We evaluate six GNN architectures based on GCN, SAGE, GIN, and GAT across five datasets from two distinct domains, employing four interpretability metrics: Fidelity, Stability, Consistency, and Sparsity. Our study examines how defenses against poisoning and evasion attacks, applied before and during model training, affect interpretability and highlights critical trade-offs between robustness and interpretability. The framework will be published as open source. The results reveal significant variations in interpretability depending on the chosen defense methods and model architecture characteristics. By establishing a standardized benchmark, this work provides a foundation for developing GNNs that are both robust to adversarial threats and interpretable, facilitating trust in their deployment in sensitive applications.

Robustness questions the interpretability of graph neural networks: what to do?

TL;DR

This work addresses the intertwined challenges of robustness and interpretability in graph neural networks under adversarial threats. It introduces a comprehensive, open-source benchmark that evaluates six GNN architectures across five datasets using four interpretability metrics (Fidelity, Stability, Consistency, Sparsity) while applying several defenses against poisoning and evasion attacks. The study finds that Fidelity and Consistency are relatively stable across settings, whereas Sparsity and Stability are more sensitive to domain, architecture, and defense choices; most defenses tend to improve interpretability, though effects vary by method and model. By revealing nuanced interactions between robustness and interpretability, the paper provides a practical foundation for designing GNNs that are both trustworthy and resilient, and it offers a reusable framework for future research.

Abstract

Graph Neural Networks (GNNs) have become a cornerstone in graph-based data analysis, with applications in diverse domains such as bioinformatics, social networks, and recommendation systems. However, the interplay between model interpretability and robustness remains poorly understood, especially under adversarial scenarios like poisoning and evasion attacks. This paper presents a comprehensive benchmark to systematically analyze the impact of various factors on the interpretability of GNNs, including the influence of robustness-enhancing defense mechanisms. We evaluate six GNN architectures based on GCN, SAGE, GIN, and GAT across five datasets from two distinct domains, employing four interpretability metrics: Fidelity, Stability, Consistency, and Sparsity. Our study examines how defenses against poisoning and evasion attacks, applied before and during model training, affect interpretability and highlights critical trade-offs between robustness and interpretability. The framework will be published as open source. The results reveal significant variations in interpretability depending on the chosen defense methods and model architecture characteristics. By establishing a standardized benchmark, this work provides a foundation for developing GNNs that are both robust to adversarial threats and interpretable, facilitating trust in their deployment in sensitive applications.
Paper Structure (37 sections, 18 equations, 1 figure, 13 tables)

This paper contains 37 sections, 18 equations, 1 figure, 13 tables.

Table of Contents

  1. Introduction
  2. Review
  3. Approaches to machine learning models attacks and defense
  4. Approaches to machine learning models interpretability
  5. Evaluation of interpretability
  6. Interactions among defense and interpretation methods in pipeline
  7. Methods
  8. Problem statement
  9. Metrics
  10. Experiments
  11. Setup of Experiments
  12. Datasets
  13. The architecture of GNNs models
  14. Training parameters
  15. Defense and interpretation methods
  16. ...and 22 more sections

Figures (1)

  • Figure 1: Overall Benchmark Pipeline. One of the available datasets $G_i$ can be selected, along with an attack method $\text{Att}$ and a defense method $D$. The next stage involves training a GNN $f$ on the selected graph $G_i$ while applying the chosen attack and defense methods. The trained model $f^{\text{trained}}$ and the dataset $G_i$ are then passed to the interpretation module, where one of the available interpretation methods can be applied to generate a mask $M(G_i)$ that highlights the important subgraph. Based on the generated masks, interpretability metrics can be computed. A detailed description of all definitions is provided in Section \ref{['section_methods']}.

Theorems & Definitions (2)

  • Definition 3.1: Defense Method for Graph Neural Networks
  • Definition 3.2: Interpretation Result in Graph Neural Networks