Trustworthy Graph Neural Networks: Aspects, Methods and Trends

He Zhang, Bang Wu, Xingliang Yuan, Shirui Pan, Hanghang Tong, Jian Pei

TL;DR

Graph Neural Networks (GNNs) are revolutionizing fields from drug discovery to social media, but their "black-box" nature and vulnerability to bias or attack pose significant risks. This comprehensive survey provides a roadmap for building Trustworthy GNNs, categorizing the field into six essential pillars: robustness, explainability, privacy, fairness, accountability, and environmental well-being.

What Problem Does This Paper Address?

While GNNs excel at learning from complex, relational data, most research has focused solely on maximizing accuracy. In high-stakes domains like medical diagnosis or credit scoring, a high-performing model that is easily fooled by adversarial attacks, discriminates against specific groups, or leaks private user data is not just flawed—it’s dangerous.

The authors argue that "trustworthiness" must be a first-class citizen in GNN design. They tackle the challenge of defining what makes a GNN trustworthy and provide a structured way to navigate the massive, fragmented body of research aimed at fixing these "unintentional harms."

Methods

The authors propose an open framework for Trustworthy GNNs, organizing the landscape into a fine-grained taxonomy (see Figure 1).

Fig. 1. The Open Framework of Trustworthy GNNs and Overview of Methodology Categorisation. Trustworthy GNNs are competent GNNs that incorporate aspects of trustworthiness, including robustness, explainability, privacy, fairness, accountability, well-being, and other trust-oriented aspects in the context of GNNs. For each aspect of trustworthy GNNs, we comprehensively summarise typical methods and provide methodology categorisation, as will be shown in the following sections.

Key methodological areas include:

  • Robustness: Analyzing how GNNs can be compromised by poisoning (training-time) or evasion (inference-time) attacks, and how to defend them via preprocessing, architecture optimization, or adversarial training. Figure 2 illustrates these attack vectors.
  • Explainability: Moving beyond black boxes using intrinsic methods (designing models that are interpretable by nature) or post-hoc explainers (tools like GNNExplainer that highlight which nodes or edges influenced a specific prediction).
  • Privacy: Implementing techniques like Federated Learning (FL) and Differential Privacy (DP) to protect sensitive graph structures and node attributes.
  • Fairness: Mitigating bias through fair representation learning or prediction enhancement, ensuring models don't discriminate based on protected attributes like race or gender.
  • Accountability & Well-being: Developing benchmarking standards for model integrity and focusing on "environmental well-being" by optimizing GNN efficiency for edge devices and reducing carbon footprints.
Fig. 2. Process of different adversarial attacks against GNNs. Training and inference without adversarial perturbations can produce a clean GNN model and correct prediction results (as shown in the second column dotted box). Meanwhile, different perturbations to the clean graph are added during either a training or inference phases to produce evasion attacks (the first column dotted box), poisoning attacks (the third column dotted box), and backdoor attacks (the fourth column dotted box), which lead to different kinds of misclassifications (“Incorrect Results” box at the bottom).

Datasets

The paper notes that while standard datasets like TUDataset and OGB are used for general performance, trustworthy GNNs require specialized benchmarks:

  • Explainability: Synthetic datasets like BA-shapes and Tree-Cycle (where ground truth motifs are known) and real-world biochemical datasets like MUTAG or BBBP.
  • Fairness: Datasets containing sensitive attributes, such as Pokec (social network), German Credit, Recidivism, and Credit Defaulter.
  • Privacy/Robustness: Often evaluated on standard citation networks (Cora, Citeseer) but under specific attack or privacy-preserving constraints.

Key Insights and Ideas

  • Graphs are Different: Trustworthiness in GNNs isn't just "standard AI safety applied to graphs." Because graph data is discrete and non-IID (nodes are interdependent), perturbations like adding/deleting an edge are fundamentally different from changing a pixel in an image.
  • The Trustworthiness Trade-off: There is a complex interplay between these pillars. For example, increasing a model's robustness or privacy often comes at the cost of "environmental well-being" due to increased computational overhead.
  • Cross-Aspect Synergy: Conversely, some aspects help each other. Fair representation learning can naturally improve privacy by making it harder to infer sensitive attributes from the model's embeddings.
  • Taxonomy of Differences: The authors provide a vital comparison (see Table II) showing how the "efficiency bottleneck" in GNNs stems from data scale and irregularity, whereas in standard AI, it's often about model scale.
Typical Research Differences: AI vs GNNs

Key Results

The survey synthesizes findings from hundreds of papers to show that:

  • Vulnerability is Real: GNNs are highly sensitive to structural perturbations; even changing a few edges can flip a classification.
  • Explainability is Maturing: Post-hoc explainers are becoming more sophisticated, but they often lack "faithfulness"—meaning the explanation might not actually reflect how the model truly works.
  • Efficiency is the New Frontier: With GNNs moving to edge devices (like autonomous vehicles), model compression and hardware accelerators are becoming essential for "trustworthy" deployment.

Limitations and Future Work

The authors identify several critical gaps:

  1. Lack of Ground Truth: Evaluating explainability is difficult because we rarely have "ground truth" explanations for real-world graphs.
  2. Scalability: Many current defense and privacy techniques are too computationally expensive for giant-scale graphs (e.g., social networks with billions of edges).
  3. Unified Ecosystem: There is a need for a "Technology Ecosystem" for Trustworthy GNNs—standardized tools, metrics, and pipelines similar to what exists for traditional machine learning.
  4. Model-Agnostic Methods: Future research should favor methods that can be "plugged in" to any GNN architecture rather than requiring custom-designed networks.

Summary updated

Abstract

Graph neural networks (GNNs) have emerged as a series of competent graph learning methods for diverse real-world scenarios, ranging from daily applications like recommendation systems and question answering to cutting-edge technologies such as drug discovery in life sciences and n-body simulation in astrophysics. However, task performance is not the only requirement for GNNs. Performance-oriented GNNs have exhibited potential adverse effects like vulnerability to adversarial attacks, unexplainable discrimination against disadvantaged groups, or excessive resource consumption in edge computing environments. To avoid these unintentional harms, it is necessary to build competent GNNs characterised by trustworthiness. To this end, we propose a comprehensive roadmap to build trustworthy GNNs from the view of the various computing technologies involved. In this survey, we introduce basic concepts and comprehensively summarise existing efforts for trustworthy GNNs from six aspects, including robustness, explainability, privacy, fairness, accountability, and environmental well-being. Additionally, we highlight the intricate cross-aspect relations between the above six aspects of trustworthy GNNs. Finally, we present a thorough overview of trending directions for facilitating the research and industrialisation of trustworthy GNNs.
BibTeX
@inproceedings{zhang2022trustworthy,
  author = {He Zhang and Bang Wu and Xingliang Yuan and Shirui Pan and Hanghang Tong and Jian Pei},
  title = {{Trustworthy Graph Neural Networks: Aspects, Methods and Trends}},
  booktitle = {Proc. IEEE},
  year = {2022},
  url = {https://openreview.net/forum?id=JTXgQOy9Vl},
}
Figure