PPGSL

Code for "Learning-based Privacy-Preserving Graph Publishing Against Sensitive Link Inference Attacks"

Requirements

• matplotlib==3.8.0
• numpy==2.0.1
• pandas==2.2.2
• scikit_learn==1.2.2
• scipy==1.14.0
• torch==2.2.2
• torch_geometric==2.5.3

Install dependencies:

pip install -r requirements.txt

File Directory

.
├─ README.md
├─ Parameter_settings_and_experimental_results_of_PPGSL.xlsx  # Recording the detailed parameter settings and experimental results of PPGSL ($k=0$) and PPGSL ($k=1$) in the paper.
├─ requirements.txt
├─ results.txt  # Recording the evaluation resluts, automatically generated after running run.py.
├─ run.py  # Running PPGSL framework, including the training and inference phase, outputing the evaluation resluts.
├─ run_attack.py  # Conducting various inference attacks (CN, AA, RA, GAE, N2V+sim, N2V+ML, GAE+sim, GAE+ML) on the published graph.
│
├─ data  # Downloaded datasets.
│  ├─ Cora
│  └─ LastFMAsia
│
├─ log_Cora
│  ├─ Cora_PPGSLsp_500_0.01_1.0.log  # Running log files. Parameter setting: dataset: Cora, method:PPGSL_sp, epoch=500, $\alpha$=0.01, $k$=1.0.
│  ├─ Cora_PPGSLsp_500_0.01_1.0_1.pt  # The learned privacy-preserving graph for publication in the 1-st repetition of the experiment.
│  ├─ Cora_PPGSLsp_500_0.01_1.0_2.pt  # The learned graph in the 2-nd repetition.
│  ├─ Cora_PPGSLsp_500_0.01_1.0_3.pt  # The learned graph in the 3-rd repetition.
│  ├─ Cora_PPGSLsp_500_0.01_1.0_4.pt  # The learned graph in the 4-th repetition.
│  └─ Cora_PPGSLsp_500_0.01_1.0_5.pt  # The learned graph in the 5-th repetition.
│
└─ PPGSL  # Implementation of PPGSL framework.
   ├─ attack.py
   ├─ dataset.py
   ├─ evaluate.py
   ├─ model.py
   ├─ origin.py
   ├─ params.json
   ├─ ppgsl_fgp.py
   ├─ ppgsl_sparse.py
   ├─ utils.py
   └─ __init__.py

Arguments Description

run.py

Name	Default value	Description
method	PPGSLsp	'none': no protection; 'PPGSLfgp': use PPGSL-FGP to conduct privacy protection; 'PPGSLsp': use PPGSL-sparse to conduct privacy protection.
dataset	Cora	Datase name, can be chosen from: {PolBlogs, LastFMAsia, DeezerEurope, Cora, CiteSeer, PubMed}.
neg_ratio	1.0	The value of sampling factor $k$ for PPGSL-sparse.
alpha	0.01	The value of parameter $\alpha$ to control the trade-off between the privacy protection effect and data utility level.
T	COS	The prediction head in the training phase. COS: cosine similarity; IP: inner product; MLP: MLP predictor.
A	COS	The prediction head of the inference attack. COS: cosine similarity; IP: inner product; MLP: MLP predictor.
seed	42	Random seed.
times	5	The number of repetitions of the experiment.
draw	False	Whether or not to plot the variations in loss and privacy/utility evaluation results as the number of epochs increases in the training process of graph learner.
learning_rate	None	Learning rate of the graph learner.
hidden_dim	None	Dimension of the hidden layers of the GNN encoder of surrogate attack model.
save_path	None	The file name to save log and output results.
device	None	Running environment, cpu or cuda.
epoch	None	Training epoch of the graph learner.
gae_epoch	None	Training epoch of the GNN encoder of surrogate attack model, also the training epoch of the GNN encoder when evaluating the link prediction utility task.
gcn_epoch	None	Training epoch of the GNN encoder when evaluating the node classification utility task.
test_interval	None	Interval of graph learner training epochs for evaluation.
update_interval	None	The value of sampling factor $\mu$ for PPGSL-skip. Interval of graph learner training epochs for updating surrogate attack model.

Note: If some arguments are set to "None", they will be assigned default values which are used in our experiments automatically. The specific assignment can be seen in PPGSL/params.json and PPGSL/utils.py.

run_attack.py

Name	Default value	Description
path_name	Cora_PPGSLsp_500_0.01_1.0	The file name of the learned graph for evaluation.
method	PPGSLsp	'none': no protection; 'PPGSLfgp': use PPGSL-FGP to conduct privacy protection; 'PPGSLsp': use PPGSL-sparse to conduct privacy protection.
dataset	Cora	Datase name, can be chosen from: {PolBlogs, LastFMAsia, DeezerEurope, Cora, CiteSeer, PubMed}.
seed	42	Random seed.
times	5	The number of repetitions of the experiment.
device	None	Running environment, cpu or cuda.
gae_epoch	None	Training epoch of the GNN encoder of the GAE-based attack model.
n2v_epoch	None	Training epoch of the node2vec model of the N2V-based attack model.
seal_epoch	None	Training epoch of the SEAL attack model.
seal_lr	None	Learning rate of the SEAL attack model.

Note: If some arguments are set to "None", they will be assigned default values which are used in our experiments automatically. The specific assignment can be seen in PPGSL/params.json and PPGSL/utils.py.

Running Examples

• Example 1: Run PPGSL-sparse ($k$=1) on Cora dataset.

  python run.py --method=PPGSLsp --dataset=Cora --neg_ratio=1.0

• Example 2: Evaluation the privacy protection effects of the learned graph 'Cora_PPGSLsp_500_0.01_1.0' against various inference attack methods.

  python run_attack.py --method=PPGSLsp --dataset=Cora --path_name=Cora_PPGSLsp_500_0.01_1.0