Deceptive Fairness Attacks on Graphs via Meta Learning

Authors: Jian Kang, Yinglong Xia, Ross Maciejewski, Jiebo Luo, Hanghang Tong

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experimental evaluations on real-world datasets in the task of semi-supervised node classification. The experimental results demonstrate that FATE could amplify the bias of graph neural networks with or without fairness consideration while maintaining the utility on the downstream task.
Researcher Affiliation Collaboration 1University of Rochester, {jian.kang@, jluo@cs.}rochester.edu 2Meta, yxia@meta.com 3Arizona State University, rmacieje@asu.edu 4University of Illinois Urbana-Champaign, htong@illinois.edu
Pseudocode Yes Appendix B presents the pseudocode of FATE. Algorithm 1 summarizes the detailed steps on fairness attack with FATE.
Open Source Code Yes Code can be found at the following repository: https://github.com/jiank2/FATE. ... the code will be publicly released under CC-BY-NC-ND license upon publication
Open Datasets Yes We use three widely-used benchmark datasets for fair graph learning: Pokec-z, Pokec-n and Bail.
Dataset Splits Yes For each dataset, we use a fixed random seed to split the dataset into training, validation and test sets with the split ratio being 50%, 25%, and 25%, respectively.
Hardware Specification Yes All experiments are performed on a Linux server with 2 Intel Xeon Gold 6240R CPUs and 4 Nvidia Tesla V100 SXM2 GPUs, each of which has 32 GB memory.
Software Dependencies Yes All codes are programmed in Python 3.8.13 and Py Torch 1.12.1.
Experiment Setup Yes Surrogate model training. ...The surrogate GCN in FA-GNN is trained for 500 epochs with a learning rate 1e-2, weight decay 5e-4, and dropout rate 0.5. For FATE, we use a 2-layer linear GCN...trained for 500 epochs with a learning rate 1e-2, weight decay 5e-4, and dropout rate 0.5. Training the victim model. ...The hidden dimension, learning rate, weight decay and dropout rate of GCN and Fair GNN are set to 128, 1e-3, 1e-5 and 0.5, respectively.