acceptodds
Under review as a conference paper at ICLR 2027

Cross-Encoder Fairness Degradation in Text-Attributed Graphs via Adversarial Node Injection

Abstract

Graph Neural Networks (GNNs) have emerged as the dominant paradigm for modeling graph-structured data. Despite their wide adoption, recent studies reveal that GNNs are vulnerable to adversarial perturbations, particularly fairness attacks. These attacks can induce discriminatory outcomes and cause severe societal consequences in human-centered applications. Consequently, existing research has investigated fairness attacks targeting GNNs on traditional graphs to expose potential risks and ultimately guide the design of more robust systems. However, to the best of our knowledge, no existing fairness attack methods target GNNs processing Text-Attributed Graphs (TAGs), even though TAGs are ubiquitous and critically important in semantically rich domains like social networks. As a result, the community currently lacks a fundamental understanding of the fairness vulnerabilities of GNNs when processing TAGs. To bridge this gap, we propose a Universal Fairness Poisoning Attack against GNNs on TAG (UFPAT), which induces fairness-bias in the victim GNN by injecting malicious nodes. UFPAT achieves the first attack of generalizing across different text encoders. It leverages Large Language Models to refine the textual features of malicious nodes, ensuring high readability for real-world feasibility. We benchmark UFPAT against SOTA methods on three real-world TAG datasets. Results show that UFPAT is the only method capable of launching fairness poisoning attacks targeting GNNs on TAGs while generating readable text. It generalizes cross-text-encoders and attacks both standard and fairness-aware GNNs. This work reveals key fairness vulnerabilities in GNNs for TAGs and offers insights for improving robustness.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Graph Neural Network, Fairness, Adversarial Attack

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_881wOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to design an adversarial attack to compromise fairness in node classification task for text-attributed graphs (TAGs) while at the same time preserving utility. Such fairness adversarial attacks in the existing literature mostly consider traditional graphs. To this end the authors propose Universal Fairness Poisoning Attack (USPAT), which attacks fairness by injecting malicious nodes. In the first stage of the attack, LLMs are leveraged to obtain textual features of the malicious nodes, followed by a second stage where edges are introduced. USPAT is deployed on three datasets/tasks, where it achieves better performance in terms of attack success as well as preserving utility.

Critical strengths

The paper is easy to follow and presents the ideas clearly.

It introduces a novel attack strategy, which is computationally efficient and is demonstrated to be successful in attacking fairness in TAGs.

Evaluation is performed on three datasets to demonstrate its effectiveness.

Critical weaknesses

I honestly struggle to understand the specific importance of the text-attribution setup. I see that the baselines are not as effective as USPAT, but the comparison is a bit unfair. The baseline attack strategies only poison the graph structure and not the associate text. So, the improvement in attack success rate of USPAT is not surprising.

The attack strategy is also a bit unrealistic. Consider a real-world network and an adversary. I would imagine the adversary to have an overview of the network (perhaps some text shared in one's profile) and only control the nodes that he introduces. When introducing edges, he can only do so between nodes he introduces but cannot make the nodes not in his control to form new edges. Assuming that an ordinary node need not form an edge with an adversarial node, would the attack still be equally effective?

Another important requirement for the success of a poisoning attack is its imperceptibility. It might be that the edges are introduced in such a way that it is straightforward to distinguish poisoned nodes from the real ones. The authors should consider developing tests to investigate how imperceptible are the poisoned nodes from normal nodes.

In Appendix J.2, the authors demonstrate how budget influences the effectiveness of the attack. The budget is measured in terms of the number of poisoned nodes. This investigation could be extended to also include other baselines. Given a fixed budget, how effective are the different baselines? From the settings, it is unclear if all the baseline attack strategies had similar budget.

Questions for authors

What was the assigned budget for each baseline?

How difficult is it to distinguish poisoned nodes from normal ones?

What sort of text is the LLM generating for the poisoned nodes? Are these realistic? I mean, the generated text might have semantic differences, making them easier to detect.

Reviewer_j6yYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work focuses on adversarial poisoning attacks of GNNs applied to TAGs, studying how fairness can be degraded. The authors argue that the intersection of TAG processing, adversarial attacks to fairness, and graph learning are understudied. To address this, the authors propose UFPAT, a two stage method that first uses an LLM to generate text for a set of nodes injected into the graph, and then optimizes how these nodes are connected to existing nodes in the graph to decrease fairness while aiming to maintain overall performance. The experiments are conducted on three main benchmarks, focusing on different standard and fairness-aware GNN architectures. Across their experiments, the authors show that fairness metrics such as DP and EO tend to degrade under their attack, with relatively small drops in accuracy. The authors also show transferability across text encoders, as they had mentioned in their motivation. Overall, I think the paper investigates an interesting vulnerability despite it being a semi-niche problem.

Critical strengths
  1. The failure mode investigated in the paper is meaningful, especially given LLM generated embeddings is becoming standard practice in graph learning. Likewise, the decoupling of overall performance and fairness is useful to highlight in these settings.

  2. The empirical results seem promising, with pretty large increases in EO and DP while maintaining relatively similar accuracies. Thus, I feel the goal of “hiding” degraded fairness via maintaining decent overall accuracy is achieved.

  3. I also think the authors offer some interesting diagnostic to understand what is injected into the graph by the LLM and what the properties of the generated text are.

Critical weaknesses
  1. One issue with the paper is the claim that the paper focuses on a “realistic and conservative threat model” simply because the GNN and text embedder are not accessible to the attacker. The method still requires a ton of access points to work: (a) a way to inject nodes into the graph, (b) a way to induce edges, (c) knowledge of other nodes in the network including their structure, original text, labels, and sensitive attributes. Although this is acknowledged, I feel like this actually needs to be addressed to make the work meaningful.

  2. The method is designed around two core pieces: (a) a pipeline where an LLM can generate realistic text that is grounded in the data, and (b) a soft optimization paradigm over the added structure to induce fairness drops. One could argue that the second piece is a well-studied technique in graph learning/graph optimization, and thus is not necessarily novel. This means that the novelty hinges more on (a) and the use of TAGs. However, the appendix shows that ablating the LLM writer and adding random text performs comparably. I understand the authors acknowledge this, but I don’t feel convinced of the argument: “Moreover, the attack remains effective when the LLM module is removed (PLLM). This result validates a key innovation of UFPAT: its attack effectiveness does not depend on the textual content of the injected nodes. The LLM is introduced primarily to generate human-readable features for malicious nodes, thereby enhancing their stealthiness.” How do the authors justify the novelty here? I also think a more rigorous analysis of the text generation, such as what the text actually looks like, other strategies for generation, etc. is important to claim it is a necessary part of the process.

  3. Many of the motivating claims appear much stronger than what is actually shown in the paper. As an example, the authors claim there is a cross-encoder transferability issue for current attacks, but don’t ever show that this occurs in practice. Can the authors more carefully show that cross-encoder transferability is actually an issue? Likewise, given the model doesn’t even seem to be sensitive to the text a node has, couldn’t this explain the transferability? More clearly showing that the text is effectively used and meaningfully transfers would make for a stronger argument.

  4. In general, the paper provides little insight into what the method is actually doing to the graph. The paper doesn’t articulate how many edges are added, what the structure of these added subgraphs are, what the generated text looks like, etc. Given this is a method heavy paper, more insight should be provided to provide evidence the attack is doing something meaningful. For instance, is there some particular property that the optimization is inducing on the graph that leads to the degraded fairness? Answering this can also help the authors ground their method in the literature of GNN fairness where properties, such as heterophily, have been shown to interact with fairness.

Questions for authors

Please see weaknesses.

Reviewer_gtp6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UFPAT, a fairness poisoning attack against GNNs on text-attributed graphs. It generates readable textual attributes for injected nodes and optimizes their connections through Relaxed Bernoulli parameterization, a straight-through estimator, and bi-level optimization. The attack aims to increase group fairness disparities while limiting accuracy degradation and transferring across text encoders.

Critical strengths
  • The paper studies a relevant problem: adversarial manipulation can amplify group disparities while leaving aggregate accuracy relatively unaffected. Generating readable textual attributes also improves the plausibility of injected nodes.

  • The evaluation includes real-world sensitive attributes and covers multiple standard and fairness-aware GNN architectures. The results demonstrate effective fairness degradation across the evaluated settings.

  • The paper is generally well organized. The supplementary experiments on restricted sensitive-attribute access, ablations, computational cost, and defenses provide useful additional evidence.

Critical weaknesses
  • The specific challenges of fairness attacks on TAGs require stronger justification. Readability constrains direct feature manipulation, but does not by itself establish why existing structure-oriented fairness attacks would be insufficient. Similarly, cross-encoder transferability is plausible for attacks driven primarily by graph structure. Table 6 demonstrates that UFPAT transfers across the tested encoders, but does not compare its transferability with relevant baselines. Preliminary or comparative experiments would help establish the claimed difficulty and the advantage of the proposed approach.

  • The methodological novelty appears limited. The overall pipeline combines node injection, text generation, a fairness-oriented objective, and bi-level structural optimization. The manuscript explains how its Relaxed Bernoulli/STE parameterization differs from previous edge-selection approaches, but the empirical evidence does not isolate the value of this particular choice. Comparing optimized connections with random connections establishes the importance of structural optimization, rather than the advantage of the proposed optimization mechanism over existing alternatives.

Questions for authors

Please see weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.