Self-supervised Adversarial Purification for Graph Neural Networks

Authors: Woohyun Lee, Hogun Park

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments across diverse datasets and attack scenarios demonstrate the state-of-the-art robustness of GPR-GAE, showcasing it as an independent plug-and-play purifier for GNN classifiers.
Researcher Affiliation Academia 1Department of Computer Science and Engineering, Sungkyunkwan University, Suwon, South Korea.
Pseudocode Yes E. Algorithms Algorithm 1 Training of GPR-GAE Algorithm 2 Multi-Step Purification with GPR-GAE
Open Source Code Yes Our code can be found in https://github.com/woodavid31/GPR-GAE.
Open Datasets Yes We conducted experiments on various datasets including Cora, Cora ML, Citeseer (Bojchevski & G unnemann 2018), Pubmed (Sen et al. 2008), OGB-ar Xiv (Hu et al. 2020), and Chameleon with removed duplicates (Platonov et al. 2023).
Dataset Splits Yes We use an inductive split with 20 labeled nodes per class for train and validation, a stratified test set of 10% of nodes, and the remaining nodes as unlabeled training data. For Chameleon and OGB-ar Xiv, we use their provided splits with fully labeled training sets.
Hardware Specification Yes All experiments are conducted on an NVIDIA A100 (80GB) GPU. However, it is worth noting that GPR-GAE can be trained and applied to datasets, including OGB-ar Xiv, using an NVIDIA RTX A5000 (24GB).
Software Dependencies No The paper describes the models used and their configurations (e.g., Two-layer GCN, MLP with 64 hidden units), but does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes GPR-GAE is trained using the ADAM optimizer with a learning rate of 0.01 and a weight decay of 0.0001. Training is conducted for 2000 epochs. ... When training the classifiers, a maximum of 3000 epochs is used for training, using the Adam optimizer with a learning rate of 0.01, weight decay of 0.001, and tanh Margin loss. An early stop method is used with a patience of 200 epochs.