Graph neural networks: a survey on the links between privacy and security

Faqian Guan, Tianqing Zhu, Wanlei Zhou, Kim-Kwang Raymond ChooArtif. Intell. Rev 2024 · DOI

Abstract

Graph Neural Networks (GNNs) have emerged as a powerful tool for processing relational data, yet their unique structural dependencies introduce novel vulnerabilities. This survey provides a unified taxonomy that bridges the gap between GNN security (adversarial robustness) and privacy (data confidentiality). By analyzing the intrinsic links between these two domains, the authors reveal how security breaches can facilitate privacy leaks and vice-versa, offering a comprehensive framework for future robust and private graph learning.

Introduction: Addressing the Research Gap

While standard deep learning models excel with "Euclidean" data like images or text, GNNs handle "non-Euclidean" graph data where relationships are as critical as the data points themselves. This complexity makes them significantly harder to secure. Historically, research has treated security—preventing malicious manipulation of model outputs—and privacy—preventing the theft of sensitive training data—as separate silos.

Comparison of existing literature highlighting the gap in research addressing the links between privacy, security, and GNNs.

As shown in the literature review above, most existing surveys focus on either security or privacy in isolation. This paper systematically maps the interdependencies between these fields, arguing that a security breach (e.g., an adversarial perturbation) can lead to a privacy leak, and a privacy attack (e.g., stealing model parameters) can make a security attack significantly more lethal.

The Multi-Level Nature of Graph Data

The vulnerability of GNNs stems from the multi-faceted nature of graph data. Unlike traditional models that operate on independent samples, GNNs must account for information at various granularities, from individual nodes to the global structure.

The hierarchical levels of graph data: node-level, link-level, subgraph-level, and graph-level.

Attackers can target any of these levels:

  • Node-level: Manipulating individual node features.
  • Link-level: Adding or deleting edges to alter the graph topology.
  • Subgraph-level: Targeting specific communities or clusters within the network.
  • Graph-level: Influencing the global properties of the entire dataset.

A Unified Taxonomy of Attacks and Defenses

The authors propose a multi-dimensional taxonomy to categorize the current landscape of GNN threats:

  1. Adversarial Attacks: Classified by the attacker's background knowledge (White-box vs. Black-box), the type of perturbation (Graph Modification vs. Graph Injection), and the attack stage (Training-time Poisoning vs. Inference-time Evasion).
  2. Privacy Attacks: These include Model Extraction (stealing the model itself), Membership Inference (determining if a specific node was part of the training set), and Attribute Inversion (reconstructing sensitive node features or edges).
  3. Defenses: The survey reviews Differential Privacy (DP) for data protection alongside security-focused techniques like Adversarial Training, Robust Message Passing, and Robustness Certificates.

Key Insights: Two Sides of the Same Coin

The central thesis of the survey is that security and privacy are intrinsically linked. For instance, if an attacker successfully executes a privacy attack to steal a GNN model, they transform a difficult "Black-box" environment into an easy "White-box" one, enabling far more damaging adversarial attacks.

Furthermore, the very mechanism that empowers GNNs—message passing—is also a primary vector for "infection." A single perturbed node can propagate its malicious representation to its neighbors, allowing a localized attack to compromise large sections of the graph. Interestingly, the authors note a "Robustness-Privacy Trade-off": some defense mechanisms, like adversarial training, can inadvertently increase the success rate of membership inference attacks, highlighting a conflict between model robustness and data confidentiality.

Results and Future Frontiers

Quantitative trends indicate that model extraction attacks on GNNs are remarkably effective, often achieving ~85% fidelity to the victim model. Structural perturbations (changing edges) are generally more effective and harder to detect than feature perturbations. While Adversarial Training remains common, "Robustness Certificates"—mathematical proofs of prediction stability—are emerging as the gold standard for high-stakes applications.

The survey concludes by identifying critical areas for future work, including the need for models inherently robust to edge-level manipulation, standardized defense frameworks for graph-level privacy, and the exploration of "Machine Unlearning" to remove specific node influences from trained models. Most importantly, there is an urgent need for joint security-privacy research to evaluate how defenses in one domain impact vulnerabilities in the other.

Summary updated

BibTeX
@article{guan2024graph,
  author = {Faqian Guan and Tianqing Zhu and Wanlei Zhou and Kim-Kwang Raymond Choo},
  title = {{Graph neural networks: a survey on the links between privacy and security}},
  journal = {Artif. Intell. Rev.},
  year = {2024},
  doi = {10.1007/s10462-023-10656-4},
}
Figure