HKUST-KnowComp / Persona_leakage_and_defense_in_GPT-2

Code for NAACL 2022 paper "You Don’t Know My Favorite Color: Preventing Dialogue Representations from Revealing Speakers’ Private Personas"

Persona_leakage_and_defense_in_GPT-2

Code for NAACL 2022 paper "You Don’t Know My Favorite Color: Preventing Dialogue Representations from Revealing Speakers’ Private Personas".

Environment

• tqdm
• numpy
• pandas
• torch >= 1.7.0
• transformers == 3.4.0

Data

All data files used are put in the processed_persona folder. You can find more descirptions about data in the folder.

Training

First, make sure you setup proper path of processed_persona in config.py.

To train the defender GPT-2, run

python training_dialogpt_defense.py

Then you can use the defender GPT-2 to train the attacker (or you can just train the attacker on the pre-trained GPT-2 without defense):

python training_attacker.py

Evaluation

Run:

./eval.sh

For path_A, path_B, model_A and model_B we use in code, you can simply use the same path and model for A and B (we did some experiments with 2 different models but the results were not good).