Authorship Anonymization through Differetial Privacy

This is the source code of our paper:

H. Bo, S. H. H. Ding, B. C. M. Fung, and F. Iqbal. ER-AE: Differentially private text generation for authorship anonymization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 3997-4007, online, June 2021. Association for Computational Linguistics.

Requirements:

# Python 3.6
# Tensorflow 1.9.0 resource: https://pypi.tuna.tsinghua.edu.cn/simple/tensorflow/
$ pip3 install <path>\tensorflow-1.9.0-cp36-cp36-win_amd64.whl
$ pip3 install sklearn scipy nltk

In model.py, get_ramdom_k_embed function is to generate random k tokens for loss calculation, embed_loss_fn is to calculate embedding loss value, predict_dp_two_sets is to generate differentially private text through two-set exponential mechanism.

Sample traing data:

my_dict.json

{
    "i": 0,
    "you": 1
}

id2w.txt

i
you

yelp.txt

1 sample training sentence

Disclaimer: Part of the tensorflow autoencoder code refers to the repo: https://github.com/bohaohan/finch. It's originally folked from zhedongzheng/tensorflow-nlp but that part of the code has been deleted in the original repository.