asappresearch / domain-private-transformers

domain-private-transformers

Repository for the paper Domain Private Transformers for Multi-Domain Dialog Systems, in Findings of EMNLP 2023

This merges several repos for langauge model training:

• lm_privacy
• huggingface langauge-modeling (example)
• private-transformers

Installation

This requires python 3.9.

conda create -n <name_env> python=3.9

# Get submodules
git submodule update --init --recursive

# Install other requirements (or requirements-macos.txt)
pip install -r requirements-linux-gpu.txt

# Install submodules and/or their requirements
pip uninstall private-transformers
cd private-transformers
pip install -e .
cd ..

Training a transformer on MultiDoGO dataset

# example script to train models
bash scripts/train_evaluate_multidogo_models.sh

Download MultiDomain Goal Oriented Dialogues (MultiDoGO) dataset link at ../ location (if at another path, use that path).

0. Use a redaction function to redact data.

# Train domain classifier
PYTHONPATH="." python domain_private_transformers/domain_classifier.py \
  --split train --model_type hf ---train_epochs 5 \
  --learning_rate 1e-5 --task dogo60 --domain_tag airline_media_ins60 --maximum_start_char 0 \
  --save_path model/redaction/dogo3dom_train60_hf_5epochs_1e-5_best.pt

# Apply classifier-based redaction
PYTHONPATH="." python scripts/run_redaction.py \
  --clf_model hf --clf_path model/redaction/dogo3dom_train60_hf_5epochs_1e-5_best.pt \
  --clf_threshold 0.96 --clf_ngram_range 16,16 \
  --output_folder_name redacted-clf0.96-dogo60-3dom-rob

1. Preprocess data and create splits for DOGO dataset (other default values of params are reasonable).

# This cmd loads all <domain>.tsv files in the input dir and creates split 
# folders in output_dir

# Create splits for unannotated data
PYTHONPATH="." python training_scripts/preprocess_dogo.py \
  --input_dir "../multi-domain-goal-oriented-dialogues-dataset/data/unannotated" \
  --output_dir "data/dogo/"
  --use_domain_tokens True \
  --valid_frac 0.2 \
  --test_frac 0.2

# Create splits for redacted data
PYTHONPATH="." python training_scripts/preprocess_dogo.py \
  --split_convos_info_path data/dogo60__split_convos_info.json \
  --task_str dogo \
  --input_dir ../multi-domain-goal-oriented-dialogues-dataset/data/redacted-clf0.96-dogo60-3dom-rob \
  --output_dir data/dogo-redacted-clf0.96-dogo60-3dom-rob/ \
  --use_domain_tokens True \
  --valid_frac 0.2 \
  --test_frac 0.2

2. Run the "exp_concave" redaction schedule with this redacted data. Other options are "step", "linear", "exp_convex". The script loops over a few learning rates, you can edit the script to only do one learning rate

# This cmd loads the AIRLINE and MEDIA data (from the non-redacted in data/dogo 
# and redacted in data/dogo-redacted-clf0.9)

bash training_scripts/train_pvt_redacted_schedule.sh \ 
  outputs/dogo_pvt_sch_redacted_clf-rob0.96_exp_concave__airline_media__eps8 \ 
  data/dogo \
  data/dogo-redacted-clf0.96-dogo60-3dom-rob/ \
  "dogo" \
  "AIRLINE,MEDIA" \
  "exp_concave"

3. Get test ppl on data/dogo for the trained model

bash training_scripts/eval_dogo.sh \
  outputs/eval__dogo_pvt_sch_redacted_clf-rob0.96_exp_concave__airline_media__eps8__lr5e-5 \
  data/dogo \
  "dogo" \
  "AIRLINE,MEDIA" \
  outputs/dogo_pvt_sch_redacted_clf-rob0.96_exp_concave__airline_media__eps8__lr5e-5

4. Train public model on only airline data (the name says w_redacted_token because the redacted token was added to the tokenizer. but don't worry, only non-redacted data was used. I needed to add the redacted token to all models to keep the tokenizer vocab same)

bash training_scripts/train_pub.sh \ 
  outputs/dogo_pub_w_redacted_tok__airline \
  data/dogo \
  "dogo" \
  "AIRLINE"

5. Run Membership Inference Attack (MIA) with keyword classifier on the target model using (4.) as ref model. Use AIRLINE contexts (prompts) and MEDIA domain for checking the generations. So we prompt with AIRLINE data and check if model leaks other domain info.

PYTHONPATH="." python LM_Memorization/extraction.py \
  --use_prompts \
  --task_mode "dogo" \
  --ctx_domains "AIRLINE" \
  --data_folder data/dogo \
  --redaction_model "keyword" \
  --redaction_target_domain "MEDIA" \
  --N 10 \
  --target_model outputs/dogo_pvt_sch_redacted_clf-rob0.96_exp_concave__airline_media__eps8__lr5e-5 \
  --ref_model outputs/dogo_pub_w_redacted_tok__airline__lr5e-4

PYTHONPATH="." python LM_Memorization/extraction.py \
  --use_prompts \
  --task_mode "dogo" \
  --ctx_domains "AIRLINE" \
  --data_folder data/dogo \
  --redaction_model "bert" \
  --redaction_model_path model/redaction/dogo3dom_train60_hf_5epochs_1e-5_best.pt \
  --redaction_model_threshold 0.96 \
  --redaction_model_ngram_range 16,16 \      
  --redaction_target_domain "MEDIA" \
  --N 10 \
  --target_model outputs/dogo_pvt_sch_redacted_clf-rob0.96_exp_concave__airline_media__eps8__lr5e-5 \
  --ref_model outputs/dogo_pub_w_redacted_tok__airline__lr5e-4
  
# Saves logs in logs/mia_bert/

Citation

@inproceedings{kabra2023domain,
    title={Domain Private Transformers for Multi-Domain Dialog Systems}, 
    author={Anmol Kabra and Ethan R. Elenberg},
    booktitle={Findings of the Association for Computational Linguistics: EMNLP 2023},
    year={2023},
}