GraphMI: Extracting Private Graph Data from Graph Neural Networks
Authors: Zaixi Zhang, Qi Liu, Zhenya Huang, Hao Wang, Chengqiang Lu, Chuanren Liu, Enhong Chen
IJCAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments over several public datasets demonstrate the effectiveness of our method. In this section, we present the experimental results to show the effectiveness of Graph MI. |
| Researcher Affiliation | Collaboration | Zaixi Zhang1 , Qi Liu1 , Zhenya Huang1 , Hao Wang1 , Chengqiang Lu2 , Chuanren Liu3 , Enhong Chen1 1Anhui Province Key Lab. of Big Data Analysis and Application, School of Computer Science and Technology, University of Science and Technology of China 2Alibaba Group 3The University of Tennessee Knoxville |
| Pseudocode | Yes | Algorithm 1 Graph MI Input: Target GNN model fθ ; Node label vector Y ; Node feature matrix X; Learning rate, ηt; Iterations T; Output: Reconstructed A 1: a(0) is set to zeros 2: Let t = 0 3: while t <T do 4: Gradient descent: a(t) = a(t 1) ηt Lattack(a); 5: Call Projection operation in (13) 6: end while 7: Call Graph auto-encoder module in (14) 8: Call Random sampling module. 9: return A |
| Open Source Code | Yes | 1https://github.com/zaixizhang/Graph MI |
| Open Datasets | Yes | Our graph model inversion attack method is evaluated on 7 public datasets from 4 categories. The detailed statistics of them are listed in the appendix. Citation Networks: We use Cora and Citeseer [Sen et al., 2008]. Social Networks: Polblogs [Adamic and Glance, 2005]. Air-Traffic Networks: The air-traffic networks are based on flight records from USA and Brazil [Ribeiro et al., 2017]. Chemical Networks: AIDS [Riesen and Bunke, 2008] and ENZYMES [Borgwardt et al., 2005]. |
| Dataset Splits | Yes | To train a target model, 10% randomly sampled nodes are used as the training set. All GNN models are trained for 200 epochs with an early stopping strategy based on convergence behavior and accuracy on a validation set containing 20% randomly sampled nodes. |
| Hardware Specification | Yes | All the experiments are conducted on Tesla V100 GPUs. |
| Software Dependencies | No | The paper mentions using GNN models (GCN, GAT, Graph SAGE) but does not provide specific version numbers for these models or any underlying software libraries (e.g., Python, PyTorch, TensorFlow) used in the experiments. |
| Experiment Setup | Yes | Parameter Settings: In experiments, we set α = 0.001, β = 0.0001, ηt = 0.1 and T = 100 as the default setting. We show how to find optimal values for hyper-parameters in the following section. |