Repository for Local Differential Privacy in Graph Neural Networks: a Reconstruction Approach
The main code to train/test is in main.py.
The default experiment setting is 5 runs with different random seeds.
For the semi-synthetic datasets, the pickle files are available in the datasets folder.
Code for pre-processing these datasets are in data.py.
All other datasets are automatically downloaded.
Use the following command to train/test the model
python main.py --seed=123 --dataset=citeseer --cols_to_group=70 --model=sage --x_eps=1 --m=10 --y_eps=1 --xhops=16 --yhops=16 --num_clusters=16 --alpha=0.1
| Argument | Description |
|---|---|
| dataset | Name of dataset (citeseer, cora, dblp, facebook, german, student) |
| cols_to_group | No. of feature columns to group to reduce feature matrix sparsity |
| model | GNN architecture (sage, gat, gcn) |
| Argument | Description |
|---|---|
| x_eps | Privacy budget for one feature |
| m | No. of features to sample for GRR_FS |
| y_eps | Label privacy budget |
| Argument | Description |
|---|---|
| x_hops | No. of hops for feature propagation during reconstruction |
| y_hops | No. of hops for label propagation during reconstruction |
| num_clusters | No. of clusters for graph partitioning for LLP loss |
| alpha | Hyperparameter to control influence of LLP loss |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent