Fake Document Infilling (FDI)

This repository contains the essential code for the paper Controllable Fake Document Infilling for Cyber Deception (Findings of EMNLP 2022).

FDI is a controllable text-infilling model to generate realisitc fake copies of critical documents with moderate modification to protect the essential information and deceive adversaries.

Folder

• FDI: Proposed FDI inference pipeline
• ILM: Text infilling model implementation modified from [1]
• WE_FORGE: Reproduction of baseline [2]
• data: Our experimented datasets

Quick Start

• Create training datasets with random masking.

  cd ILM
  sh create_datasets.sh
  

• Train a general text-infilling model.

  • See sample code in ILM/training_script.txt

• Inference via controllable masking.

  • See sample code in FDI/inference_demo.ipynb

Evaluation details

• Check our designed quiz, evaluated samples, and results.

Reference

[1] Enabling language models to fill in the blanks. https://github.com/chrisdonahue/ilm

[2] Abdibayev, Almas, et al. "Using Word Embeddings to Deter Intellectual Property Theft through Automated Generation of Fake Documents." ACM Transactions on Management Information Systems (TMIS) 12.2 (2021): 1-22.

Citation

If you find this repo useful in your research, please consider citing:

  @inproceedings{hu2022controllable,
    title={Controllable Fake Document Infilling for Cyber Deception},
    author={Hu, Yibo and Lin, Yu and Parolin, Erick Skorupa and Khan, Latifur and Hamlen, Kevin},
    booktitle={Findings of the Association for Computational Linguistics: EMNLP 2022},
    pages={6505--6519},
    year={2022}
  }