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

TL;DR

This paper introduces a comprehensive benchmark to analyze how robustness defenses impact the interpretability of various GNN architectures across multiple datasets, revealing critical trade-offs and variability in interpretability under adversarial scenarios.

Contribution

It provides the first systematic benchmark evaluating the interplay between robustness defenses and interpretability in GNNs, covering multiple models, datasets, and interpretability metrics.

Findings

Interpretability varies significantly with defense methods and model architecture.

Robustness defenses can both improve and impair interpretability depending on the context.

The benchmark facilitates future development of GNNs that balance robustness and interpretability.

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.