Federated Learning on Knowledge Graphs

PyTorch code that accompanies the paper Efficient Federated Learning on Knowledge Graphs via Privacy-preserving Relation Embedding Aggregation. Our work contains Knowledge Graph Reconstruction Attack and the federated version of the following Knowledge Graph algorithms: TransE, RotatE, ComplEx, DistMult, NoGE and KB-GAT.

Prepare Dataset:

There is a data preprocessing file dataset.ipynb which allows split triples randomly. We does not provide original data that can be downloaed online, but we provide a federated version of DDB14 in folder --Fed_data.

Run Experiments:

There is script exp.sh running all federated experiments with non-gnn models, while the codes of Fed-NoGE are in folder --NoGE. Some examples of reconstruction attack are shown in rec_attack_fedr_fb15k_TransE.ipynb as well as in the folder --FedE/Rec_Attack/...

Citation

Please cite our paper if you find this code useful for your research. For any clarification, comments, or suggestions please create an issue or contact kaz321@lehigh.edu

@article{fedr,
      title={Efficient Federated Learning on Knowledge Graphs via Privacy-preserving Relation Embedding Aggregation}, 
      author={Zhang, Kai and Wang, Yu and Wang, Hongyi and Huang, Lifu and Yang, Carl and Chen, Xun and Sun, Lichao},
      journal={Findings of EMNLP},
      year={2022},
}

Acknowledgement

• FedE
• GNN-NoGE
• kg-reeval