Awesome

Related Publications

1. Poisoning Knowledge Graph Embeddings via Relation Inference Patterns (ACL-IJCNLP 2021) - <a href="https://aclanthology.org/2021.acl-long.147/">Paper</a>, <a href="https://github.com/PeruBhardwaj/InferenceAttack">Codebase</a>
2. Adversarial Robustness of Representation Learning for Knowledge Graphs (PhD Thesis) - <a href="http://www.tara.tcd.ie/bitstream/handle/2262/101176/Thesis_PeruBhardwaj.pdf">Link</a>

Update -

The initial codebase had a mistake in the implementation of Gradient Rollback. This has been corrected, and updated values are reported in Chapter 4 of the thesis <a href="http://www.tara.tcd.ie/bitstream/handle/2262/101176/Thesis_PeruBhardwaj.pdf">here</a>.

Overview

The figure illustrates adversarial attacks against KGE models for fraud detection. The knowledge graph consists of two types of entities - Person and BankAccount. The missing target triple to predict is (Sam, allied_with, Joe). Original KGE model predicts this triple as True, i.e. assigns it a higher score relative to synthetic negative triples. But a malicious attacker uses the instance attribution methods to either (a) delete an adversarial triple or (b) add an adversarial triple. Now, the KGE model predicts the missing target triple as False.

The attacker uses the instance attribution methods to identify the training triples that are most influential for model's prediciton on the target triple. These influential triples are used as adversarial deletions. Using the influential triple, the attacker further selects adversarial additions by replacing one of the two entities of the influential triple with the most dissimilar entity in the embedding space. For example, if the attacker identifies that (Sam, deposits_to, Suspicious_Account) is the most influential triple for predicting (Sam, allied_with, Joe), then they can add (Sam, deposits_to, Non_Suspicious_Account) to reduce the influence of the influential triple.

Reproducing the results

Setup

• python = 3.8.5
• pytorch = 1.4.0
• numpy = 1.19.1
• jupyter = 1.0.0
• pandas = 1.1.0
• matplotlib = 3.2.2
• scikit-learn = 0.23.2
• seaborn = 0.11.0

Experiments reported in the paper were run in the conda environment attribution_attack.yml.

Steps

• The codebase and the bash scripts used for experiments are in KGEAttack.
• To preprocess the original dataset, use the bash script preprocess.sh.
• For each model-dataset combination, there is a bash script to train the original model, generate attacks from baselines and proposed attacks; and train poisoned model. These scripts are named as model-dataset.sh.
• The instructions in these scripts are grouped together under the echo statements which indicate what they do.
• The commandline argument --reproduce-results uses the hyperparameters that were used for the experiments reported in the paper. These hyperparameter values can be inspected in the function set_hyperparams() in utils.py.
• To reproduce the results, specific instructions from the bash scripts can be run on commandline or the full script can be run.
• All experiments in the paper were run on a shared HPC cluster that had Nvidia RTX 2080ti, Tesla K40 and V100 GPUs.

References

Parts of this codebase are based on the code from following repositories

• ConvE
• Influence Functions Analysis
• Influence Functions
• CRIAGE
• Gradient Rollback
• KGC Re-evalaution
• ComplEx-N3

Citation

@inproceedings{bhardwaj-etal-2021-adversarial,
    title = "Adversarial Attacks on Knowledge Graph Embeddings via Instance Attribution Methods",
    author = "Bhardwaj, Peru  and
      Kelleher, John  and
      Costabello, Luca  and
      O{'}Sullivan, Declan",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.648",
    pages = "8225--8239",
    }