This repository implements the paper Robust Graph Convolutional Networks Against Adversarial Attacks.
The following attacks are implemented and tested:
- Random attack: generate fake edges
- RL-S2V: adversarial attack based on reinforcement learning
- NETTACK: adversarial perturbations for direct and influence attacks
The following baselines are tested:
The random-attack.ipynb notebook contains code for training and evaluating the GCN and RGCN models under a non-targeted random attack. random-attack-plot.ipynb contains logic to generate the plot of noise ratio to accuracy between GCN and RGCN during a non-targeted random attack.
The plot below shows the results of the RGCN model compared to the GCN model under a non-targeted random attack from a ratio of noise to clean edges from 0 to 1 in 0.1 increments. Each increment is generated as the mean value of test accuracies over 10 training runs for each model.
TODO
TODO
TODO