Project analyzing effectiveness of adversarial attack methods on Graph Neural Networks and defense methods. Two papers are currently considered in this project. Nettatack and GCNJaccard. All experiments are performed with implementations of these methods available at Deeprobust- which is a pytorch adversarial library for attack and defense methods on images and graphs. More details are available in the report.
Install DeepRobust library as
git clone https://github.com/DSE-MSU/DeepRobust.git
python DeepRobust/setup.py install
Then checkout this project's code as
git clone https://github.com/viz27/Attack_on_Graphs.git
cd Attack_on_Graphs
Run as given below, specifying the dataset and Jaccar Score Threshold.
python GCN_vs_GCNJaccar.py --dataset cora --threshold 0.01
| Threshold | Edges Removed(Citeseer) | Accuracy(Citeseer) | Edges Removed(Cora) | Accuracy(Cora) |
|---|---|---|---|---|
| 0.00 | 0 | 0.7198 | 0 | 0.8234 |
| 0.01 | 96 | 0.7216 | 548 | 0.8234 |
| 0.02 | 360 | 0.7174 | 548 | 0.8234 |
| 0.03 | 483 | 0.7186 | 1015 | 0.8104 |
| 0.04 | 817 | 0.718 | 1203 | 0.7953 |
| 0.05 | 1085 | 0.7079 | 1561 | 0.7741 |
| 0.06 | 1385 | 0.715 | 2046 | 0.7596 |
| 0.07 | 1658 | 0.7139 | 2248 | 0.747 |
| 0.08 | 1912 | 0.6991 | 2692 | 0.7339 |
| 0.09 | 2121 | 0.6985 | 3032 | 0.7208 |
| 0.10 | 2284 | 0.6949 | 3253 | 0.7133 |
Run as given below, specifying the dataset and Jaccar Score Threshold.
python nettack_and_defense.py --dataset cora --threshold 0.04
| Method | Accuracy(Citeseer) | Accuracy(Cora) |
|---|---|---|
| NETTACK | 0.1 | 0.15 |
| GCNJaccard(0.01) | 0.3 | 0.3 |
| GCNJaccard(0.02) | 0.525 | 0.3 |
| GCNJaccard(0.03) | 0.675 | 0.4 |
| GCNJaccard(0.04) | 0.7 | 0.45 |
| GCNJaccard(0.05) | 0.65 | 0.55 |
| GCNJaccard(0.10) | 0.775 | 0.625 |