In computer vision or natural language processing, adversarial examples are often highly specific to a model's exact parameters. A perturbation crafted to fool a specific ResNet architecture might completely fail against a VGG network because the attack exploits fine-grained, pixel-level decision boundaries unique to that model's weights.
In graph-structured data, however, a fascinating and troubling phenomenon emerges: adversarial attacks exhibit remarkable cross-architecture transferability. An adversarial perturbation optimized against a simple, linear surrogate model can effortlessly degrade the performance of a highly complex, non-linear Graph Neural Network (GNN), a Graph Transformer, or even a traditional, non-neural clustering heuristic.
To understand why graph-based attacks transfer so robustly across entirely different learning paradigms, we have to look past the weights of individual models and focus on the geometry of the data itself.
Topology Over Architecture
The fundamental reason behind cross-architecture transferability is that graph vulnerabilities are primarily rooted in the graph topology, rather than the specific mechanics of a neural network. When an adversary manipulates a graph—such as by adding or deleting a small number of edges—they are not just shifting a few isolated data points; they are altering the underlying structure of the entire data distribution.
As Wang et al. demonstrated, these structural perturbations exploit vulnerabilities inherent to the graph's geometry. Because different relational learning models and node embedding techniques must all operate on this same geometric reality, an attack that successfully distorts the graph's topology will naturally mislead almost any model attempting to parse it. The vulnerability is a property of the network's graph structure, making it largely architecture-agnostic.
The Shared Mathematical Pipeline
Underneath their diverse architectural surfaces, most graph models share a common mathematical foundation. Whether a model is a simple label propagation algorithm or a deep convolutional network, it generally relies on two core concepts: neighborhood aggregation and spectral filtering. Attackers exploit these shared mechanisms to achieve high transferability.
Neighborhood Aggregation
GNNs rely on message-passing mechanisms that aggregate information from local neighborhoods. Zügner et al. showed that attacks designed against simplified linear models transfer effectively to complex non-linear GNNs because both architectures rely on the same basic principles of graph information aggregation.
Similarly, Liu et al. exploited the mathematical equivalence between GNNs and simple Label Propagation to execute highly transferable label poisoning attacks. Because the core vulnerability stems from how information flows through local neighborhoods, the attack remains effective without needing any knowledge of the target model's internal parameters.
Spectral Filtering
From a signal processing perspective, many graph embedding models act as graph signal filters. Chang et al. demonstrated that by perturbing a graph's spectral properties rather than specific model weights, an attacker can degrade performance across multiple model types simultaneously.
This shared mathematical core means that an attacker does not even need to query a victim model to break it. As Xu et al. showed, a "strict black-box" attack can target the fundamental mathematical filters common to almost all GNNs, bypassing the need for architectural knowledge entirely.
[Adversarial Perturbation (Edge Addition/Deletion)]
│
▼
[Altered Graph Spectral Properties]
│
┌──────────────┴──────────────┐
▼ ▼
[Model A: Linear GCN] [Model B: Deep Graph Transformer]
(Performance Drops) (Performance Drops)
Bridging the Gap: From Surrogates to Heuristics
The reach of cross-architecture transferability extends far beyond transferring attacks from one neural network to another; it bridges the gap between deep learning and classical, non-neural graph algorithms.
For instance, Trappolini demonstrated that adversarial perturbations optimized against specific GNNs for link prediction transfer successfully to traditional, non-neural heuristics like Jaccard similarity or PageRank.
A similar pattern occurs in community detection. The authors of CD-ATTACK and Bernini et al. found that adversarial strategies designed to evade a specific surrogate GNN or community detection method generalize robustly to entirely different paradigms, such as modularity-based clustering, Louvain, or WalkTrap. Because these diverse algorithms all rely on similar fundamental structural features to define community boundaries, a topological perturbation that dissolves a community boundary for a surrogate model dissolves it for the others as well.
Generalizing Across Domains and Tasks
Because these vulnerabilities are structural rather than model-dependent, attackers can learn generalized strategies that transcend specific datasets, representations, or domains.
Instead of optimizing for a specific sequence of actions on a single graph, Zhang et al. developed a method that learns a global importance ranking across all possible graph modifications. This representational choice allows the attack to identify fundamental structural weaknesses that generalize across different datasets and model architectures.
Taking this a step further, Zhu et al. demonstrated that manipulating raw social interactions (such as user shares and engagements) acts as a universal proxy. Because different graph architectures are derived from the same underlying social data, perturbations at the raw data level transfer successfully between models that use entirely different graph representations, such as bipartite graphs or propagation trees.
Defensive Implications: Data-Centric Immunization
The reality of cross-architecture transferability turns traditional defense strategies on their head. If the vulnerability is inherent to the graph topology, then model-specific architectural tweaks or standard adversarial training are often insufficient.
This insight has led researchers to focus on data-centric defenses. Tao et al. and Shen et al. introduced the concept of graph adversarial immunization, which proactively "vaccinates" a small fraction of critical graph structures or edges to guarantee they cannot be modified by an attacker.
Crucially, just as attacks transfer across models because they target fundamental structural properties, these defensive measures exhibit high transferability as well. Once critical connections are secured against a surrogate attack model, the resulting structural stability protects the graph against a wide variety of unseen, sophisticated attack vectors across different architectures.
Key Foundational Work
Attacking Graph-based Classification via Manipulating the Graph Structure (2019) — contributed 1 concept elements to this direction.
Adversarial Attacks on Graph Neural Networks via Meta Learning (2019) — contributed 1 concept elements to this direction.
Adversarial Attacks on Neural Networks for Graph Data (2019) — contributed 1 concept elements to this direction.