Adversarial Attacks on Fairness of Graph Neural Networks
Binchi Zhang, Yushun Dong, Chen Chen, Yada Zhu, Minnan Luo, Jundong Li
TL;DR
This paper introduces G-FairAttack, a framework for adversarially compromising the fairness of fairness-aware GNNs without affecting their prediction accuracy, highlighting vulnerabilities in current fairness methods.
Contribution
The paper presents G-FairAttack, a novel attack framework targeting fairness in GNNs, along with a fast computation method to enhance efficiency.
Findings
G-FairAttack effectively reduces fairness in various GNN models.
The attack maintains high prediction utility, remaining unnoticeable.
Experimental results demonstrate the vulnerability of fairness-aware GNNs.
Abstract
Fairness-aware graph neural networks (GNNs) have gained a surge of attention as they can reduce the bias of predictions on any demographic group (e.g., female) in graph-based applications. Although these methods greatly improve the algorithmic fairness of GNNs, the fairness can be easily corrupted by carefully designed adversarial attacks. In this paper, we investigate the problem of adversarial attacks on fairness of GNNs and propose G-FairAttack, a general framework for attacking various types of fairness-aware GNNs in terms of fairness with an unnoticeable effect on prediction utility. In addition, we propose a fast computation technique to reduce the time complexity of G-FairAttack. The experimental study demonstrates that G-FairAttack successfully corrupts the fairness of different types of GNNs while keeping the attack unnoticeable. Our study on fairness attacks sheds light on…
Peer Reviews
Decision·ICLR 2024 poster
- The paper is well written and easy to read. - The proposed unnoticeable fairness attacks of GNNs are novel and interesting. - The theoretical analysis demonstrates that the designed surrogate loss function serves as a common upper bound for three fairness loss functions.
- Grey-box attack scenarios are relatively uncommon in real-world applications. I believe it would be more interesting if it could be extended to black-box attack settings. - In terms of the utility metrics, G-FairAttack and the baseline seem to have a relatively small difference. I believe this does not fully reflect the authors' claim of making attacks unnoticeable. In other words, the issue mentioned by the authors in the introduction, "no existing work considers unnoticeable utility change i
The introduction of G-FairAttack brings a new perspective to the understanding of adversarial attacks in the context of fairness-aware models. By uncovering vulnerabilities related to fairness, the paper contributes valuable insights that can guide the development of more robust and ethical AI systems. The paper includes extensive experiments that validate the effectiveness of the proposed attacks. This empirical evaluation strengthens the credibility of the findings and their relevance to pra
The assumptions about the attacker's knowledge might not cover all possible real-world scenarios. The gray-box setting is a middle ground, but exploring both black-box and white-box attacks could provide a fuller picture of the vulnerabilities. The performance of G-FairAttack is worse than random attack under some scenarios in Table 1, 6 and 7.
1.Relevance of Topic: The paper addresses the crucial and timely subject of fairness in GNNs, a significant area in AI research. 2.Innovative Framework: The introduction of the G-FairAttack framework, complemented by a fast computation technique, offers a novel perspective on understanding vulnerabilities in fairness-aware GNNs.
1． Unpractical attack setting. The proposed evasion attack is not practical as there is no motivation for the model owner to replace data in a transductive learning setting, as shown by equation 1. 2． Overclaimed contribution. The evidence is needed when presenting “In this way, the surrogate model trained by our surrogate loss will be close to that trained by any unknown victim loss, which is consistent with conventional attacks on model utility.” 3． Untenable theoretical analysis. The theor
Code & Models
Videos
Taxonomy
TopicsAdversarial Robustness in Machine Learning · Ethics and Social Impacts of AI