Code accompanying "DP-BART for Privatized Text Rewriting under Local Differential Privacy" paper ( https://arxiv.org/abs/2302.07636 ).

$ sudo apt-get install python3-dev

$ pip install -r requirements.txt

Experiments can be run using the scripts provided in sample_scripts. Examples are provided for the atis dataset at a privacy budget of $\varepsilon = 250$, but can be extended to other datasets and $\varepsilon$ values by changing the --dataset and --epsilon arguments, respectively.

There are three primary experimental settings in the paper: original, rewrite-no-dp and rewrite-dp.

An experiment using the original configuration can be run with downstream_experiment_original_data.sh.

Experiments using the rewrite-dp configurations can be run as follows:

• Rewriting
  • ADePT: Run with run_adept_rewrite.sh. Must specify a pre-trained model with --last_checkpoint_path, which can be prepared with the pre-training script adept_pretrain_experiment_openwebtext_glove.sh. When using GloVe embeddings, need to also specify a path to the embedding model with --embed_dir_unprocessed.
  • DP-BART-CLV: Run with run_dp_bart_clv_rewrite.sh. Must specify --last_checkpoint_path, which is the path to a pre-trained BART model.
  • DP-BART-PR: Run with run_dp_bart_pr_rewrite.sh. Need to include --pruning_index_path and --last_checkpoint_path from a pruned BART model. This can be prepared using run_dp_bart_pr_pruning.sh, which saves each checkpoint and pruned indices at every epoch.
  • DP-BART-PR+: First run additional training steps on a pruned BART model using the provided run_dp_bart_pr_plus_training.sh script for a given $\varepsilon$ privacy budget. Then run the rewriting script using this prepared model with run_dp_bart_pr_plus_rewrite.sh. Must again specify --pruning_index_path and --last_checkpoint_path from the prepared model.
• Downstream
  • The above rewriting scripts will output a rewritten dataset at the specified $\varepsilon$ privacy budget in the output directory --output_dir. Downstream experiments can then be run with the run_downstream.sh script, specifying the rewritten training dataset split with --custom_train_path, optionally a rewritten validation split with --custom_valid_path and the path to the original test set with --custom_test_path. Optionally can specify the test set with --downstream_test_data (e.g. atis, imdb, etc.).

Experiments using the rewrite-no-dp configurations can be run as above for the rewrite-dp configurations, but specifying --private False for any rewriting scripts.

Additional hyperparameters such as --batch_size, --learning_rate, and --delta can be modified in the script arguments. A full list of arguments and their description can be found in settings.py.

This repository is an extension to the framework provided at https://github.com/trusthlt/dp-rewrite. More details on running other types of configurations can be found there.