update

Author: Jiayi Luo

README.md

@@ -0,0 +1,55 @@
+# DiffSP
+
+### Code Structure Overview
+
+* **`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.
+
+### Data Download
+
+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 through an anonymous account 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.
+
+### Run The Code
+
+#### Node classification
+
+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
+```
+
+#### Graph classification
+
+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
+```
+
+#### Note
+
+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.

__pycache__/loss.cpython-311.pyc

@@ -0,0 +1 @@
+{'device_id': '0', 'dataset': 'COLLAB', 'attack': 'prbcd', 'is_train': True, 'device': device(type='cpu'), 'seed': 102, 'T_E': 9, 'name': 'COLLAB', 'data': {'batch_size': 16384, 'val_batch_size': 131072}, 'train': {'num_epochs': 10000, 'max_grad_norm': 10, 'val_every_epochs': 5, 'patient_epochs': 50}, 'mlp_X': {'hidden_t': 32, 'hidden_X': 512, 'num_mlp_layers': 2, 'dropout': 0.1}, 'gnn_E': {'hidden_t': 32, 'hidden_X': 32, 'hidden_E': 128, 'use_att': True, 'num_layers': 2, 'dropout': 0.1}, 'optimizer_X': {'lr': 0.001, 'weight_decay': 0, 'amsgrad': True}, 'optimizer_E': {'lr': 0.0003, 'weight_decay': 0, 'amsgrad': True}, 'lr_scheduler': {'factor': 0.9, 'patience': 5, 'verbose': True}, 'purify': {'k': 6, 'basic_t': 2, 'adaptive_t': 2, 'lamb': 1000.0}}

__pycache__/parameter.cpython-311.pyc

@@ -1,3 +1,9 @@
+channels:
+  - pytorch
+  - nvidia
+  - ostrokach-forge
+  - conda-forge
+  - defaults
 dependencies:
   - _libgcc_mutex=0.1
   - _openmp_mutex=4.5

__pycache__/train.cpython-311.pyc

@@ -107,7 +107,8 @@ def main(args):
     gcn = GCN(num_features, 16, num_classes).to(device)
 
     ck_path = f'./checkpoint/diffusion/{name}/n_step_E{args.T_E}'
-    # train_ddpm(args, model, train_val_dataset, device, ck_path)
+    if args.is_train:
+        train_ddpm(args, model, train_val_dataset, device, ck_path)
 
     save_gcn_path = f'./checkpoint/classifier/{name}/GCN.pth'
     gcn.load_state_dict(torch.load(save_gcn_path, map_location=device))

checkpoint/diffusion/COLLAB/n_step_E9.txt

@@ -141,7 +141,8 @@ def main(config):
     classifier = GCN(num_features, 16, num_classes).to(device)
 
     # ====================model==================== #
-    # train_ddpm(config, model, device, x, E_one_hot, ck_path)
+    if config.is_train:
+        train_ddpm(config, model, device, x, E_one_hot, ck_path)
     train_classifier(classifier, x, edge_index, y, idx_train)
 
     # ====================purify==================== #

environment.yml

@@ -10,6 +10,7 @@ def parameter_parser():
     parser.add_argument('--device_id', type=str, default='0', help='device id')
     parser.add_argument('--dataset', type=str, default='Cora', help='dataset')
     parser.add_argument('--attack', type=str, default='prbcd', help='attack')
+    parser.add_argument('--is_train', action='store_true', help='If set, train is True')
 
     args = parser.parse_args()