Research map / Graph learning / Graph self-supervised learning
Deceptive Fairness Attacks on Graphs via Meta Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FATE, a meta-learning bi-level poisoning framework that perturbs graph structure/features to deceptively amplify group fairness bias in GNN node classification while keeping utility close to benign.
Problem
Poisoning graph learning models so that fairness bias (statistical parity, individual fairness) is exacerbated while the attack remains deceptive, i.e., utility stays comparable so the attack is not detected.
Approach
Formulates fairness poisoning as a bi-level optimization solved via meta-learning, approximating the group acceptance rate with a Gaussian KDE CDF so the bias is differentiable with respect to the graph, and applies both continuous and discretized perturbations under budget constraints on graph distance and utility.
Key results
- FATE is the only compared method that consistently achieves effective and deceptive fairness attacks across datasets (NBA results in rebuttal Table 1).
- On Pokec-n, FATE changes micro F1 by 0.8% versus benign graph, while FA-GNN decreases it by 0.9%; on Pokec-z FATE increases it by 0.1% vs. FA-GNN's 2.9% decrease.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FATE instead of FA-GNN for deceptive fairness attacks | Graph node classification · Pokec-n, Pokec-z, and Bail | micro F1: 0.9% decrease for Pokec-n; 2.9% decrease for Pokec-z; 3.3% decrease for Bail; FATE: 0.8% for Pokec-n, 0.1% increase for | Helped |
| Use FA-GNN instead of FATE to increase bias | Fairness attacks on graphs · Several cases | bias | Mixed |
| Use DICE-S for statistical parity attacks instead of FATE | Statistical parity attacks · All datasets other than Bail | attack effectiveness | Hurt |
| Use DICE-S instead of FATE | Statistical parity attacks · Bail and other cases where DICE-S increases bias | bias | Hurt |
Surprising findings
- FA-GNN achieves higher statistical parity bias than FATE in several cases, but at a larger utility cost that makes it easier to detect.
Limitations
- Time and space complexity is O(n^2) due to gradients w.r.t. the dense adjacency matrix, so scaling to millions of nodes is untested.
- Only statistical parity and individual fairness are evaluated; Rawls fairness and group-specific attacks are discussed only as sketches in rebuttal.
- Fair link prediction is not supported because structural attacks may leak into test links.
Experimental setup
Datasets: Pokec-n Pokec-z Bail NBA
Benchmarks: Pokec-n Pokec-z Bail NBA
Metrics: statistical parity bias individual fairness bias micro F1
Baselines: FA-GNN DICE-S Metattack
Builds on: Metattack DICE
Closest papers by meaning
- Adversarial Attacks on Fairness of Graph Neural Networks (ICLR 2024 · similarity 0.89)
- Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.87)
- Are Your Models Still Fair? Fairness Attacks on Graph Neural Networks via Node Injections (NeurIPS 2024 · similarity 0.87)
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.86)
- Stable Fair Graph Representation Learning with Lipschitz Constraint (ICML 2025 · similarity 0.86)
- DANCE: Dual Unbiased Expansion with Group-acquired Alignment for Out-of-distribution Graph Fairness Learning (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.