Keep It Private

This is the repository for the project Keep it Private: Unsupervised Privatization of Online Text

How to Install

1. Download the model weights

pip install huggingface_hub
huggingface-cli download csbao/kip-dipper-large --local-dir models/dipper-large
huggingface-cli download csbao/kip-dipper-large dipper_cd_v130.bin --local-dir models

2. Set up the environment

./scripts/setup.sh

This will create a conda environment named kip and install all dependencies.

Keep It Private Overview

Keep it Private performs authorship transfer by performing authorship transfer using a seq2seq model that was adversarially fine-tuned via reinforcement learning using a set of rewards (Privacy, Sense, and Soundness metrics)

Input Format

The input file should be a JSONL file, one JSON object per line, with a fullText key:

{"fullText": "hi! this is the first document to privatize."}
{"fullText": "This is another text input. It can have multiple sentences..."}

Command template

$ conda activate kip
$ python src/generate.py --input_data_path ${INPUT_DATA_PATH} \
      --output_path ${OUTPUT_PATH} \
      --model_path models \
      --model_name_to_use dipper-large \
      --model_start_file ${BIN_FILE}  \ 
      --token_max_length 256 \

Example

python src/generate.py --input_data_path {JSONFILE} \
      --output_path {OUTPUT_FILE} \
      --model_path models \
      --model_name_to_use dipper-large \
      --model_start_file models/dipper_cd_v130.bin \
      --token_max_length 256

Parameters

1. input_data_path: path to the query documents to be privatized
2. output_path: file to save the privatized documents
3. model_start_file: path to trained KiP model
4. model_name_to_use: path to pre-trained base model
5. token_max_length: max cutoff length of output
6. random_seed: initialize all random seed to this value
7. lex_diversity: lexical diversity level (20, 40, or 60)
8. order_diversity: order diversity level (20, 40, or 60)