GRAP^3: A Data-free Model Extraction Attack Framework under Graph Pre-training and Prompt Learning Paradigm

Setup the environments

conda env create --file environment.yml

Conduct the model stealing attacks

python main.py --model_type SAGE --pre_dataset Arxiv --if_prompt --target_dataset Citeseer --pretrain_method GAE --seed 3407 --device 0

Explainations:

--model_type        ['SAGE', 'GCN']                    # GNN model architecture
--pre_dataset       ['Arxiv']                          # Dataset used for pretraining
--if_prompt                                            # Enable prompt mode (boolean flag)
--target_dataset    ['Cora', 'Citeseer', 'Pubmed']     # Dataset for the target model
--pretrain_method   ['DGI', 'GAE', 'supervised']       # Choose the pre-training strategy
--seed              int                                # Random seed for reproducibility
--device            int                                # GPU device ID to use