This is the implementation of Latent Adversarial Training of Graph Convolutional Networks LATGCN by Hongwei Jin and Xinhua Zhang.

• maximize the perturbation on the latent layer
• perturbation applied to all the nodes, due to the transductive learning property
• improve the accuracy
• reduce the success rate from adversarial attacks

Instal the required packages

pip install -r requirements.txt

Run the demo of nettack with vanilla GCN

python demo.py

run the demo of nettack with LATGCN

python demo.py

demo.py has a set of parameter to specify

--dataset:      choose from 
                'citeseer', 'cora', 'cora_ml', 'pubmed', 'polblogs', 'dblp'
--train_share:  specify the size for training
--reg:          a flag to toggle the LATGCN
--eta:          the norm constraint of noise applied to each node embedding
--gamma:        regularizer factor

Note that without specify the reg flag, it is simply the vanilla GCN.

Performance with perturbation on cora dataset:

Success rate with perturbation on cora dataset:

The Nettack model is implemented by Daniel Zügner
The original GCN model is implemented by Thomas N Kipf

Feel free to create issue in the repo, I will try to answer as soon as I can.

@article{jin2019latent,
  title={Latent Adversarial Training of Graph Convolutional Networks},
  author={Jin, Hongwei and Zhang, Xinhua},
  journal={ICML workshop - Learning and Reasoning with Graph-Structured Representations},
  year={2019}
}

The work is under MIT license.