checkpoint: This directory contains the trained models.data: Designated for storing the datasets used in the experiments.config: Contains the configuration files, which define the hyperparameters for training and testing.g: Includes the code for training and testing on the graph classification task.n: Includes the code for training and testing on the node classification task.model: Houses the implementation of the DiffSP model.utils: Contains the code for graph transfer entropy and node LID estimation.
To facilitate the reproduction of our experimental results, we provide the adversarial data used in our experiments.
Due to repository space limitations, the data is available at: https://huggingface.co/datasets/anonymous-random/DiffSP-data/tree/main.
Please download the data and place it in the ./data/ directory before running the experiments.
For example, we run the experiments on the Cora dataset under the PR-BCD attack.
python n/main.py --dataset Cora --attack prbcd
If you want to retrain the DiffSP, use --is_train like:
python n/main.py --dataset Cora --attack prbcd --is_train
For example, we run the experiments on the IMDB-BINARY dataset under the PR-BCD attack.
python g/main.py --dataset IMDB-BINARY --attack prbcd
If you want to retrain the DiffSP, use --is_train like:
python g/main.py --dataset IMDB-BINARY --attack prbcd --is_train
For graph classification, the classifier comprises two GCN layers, followed by a mean pooling layer and a linear layer, consistent across all models and identical to the surrogate model.
For node classification, the classifier consists of two GCN layers, consistent across all models and aligned with the surrogate model.