skywalker023 / confaide

🤫 Code and benchmark for our paper: "Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory"

🤫 ConfAIde

​ This is the official repository for our paper:
Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory

Our benchmark ConfAIde evaluates inference-time privacy implications of LLMs, in interactive settings. The benchmark has 4 tiers, and you can find the dataset/scenarios under the ./benchmark directory.

Please cite our work if you found the resources in this repository useful:

@article{confaide2023,
  author    = {Mireshghallah, Niloofar and Kim, Hyunwoo and Zhou, Xuhui  and Tsvetkov, Yulia and Sap, Maarten and Shokri, Reza and Choi, Yejin},
  title     = {Can LLMs Keep a Secret? Testing Privacy  Implications of Language Models via Contextual Integrity Theory},
  journal   = {arXiv preprint arXiv:2310.17884},
  year      = {2023},
}

​

Reproducing the results

​ First, create the conda environment by running:

conda env create -f environment.yml

​ and then activate it:

conda activate confaide

You can run the evaluation by running the following example command

python eval.py --model gpt-3.5-turbo-0613 --data-tier 1 --n-samples 10

This command will run ChatGPT on Tier 1 with 10 sampled results for each data point. You can choose Tiers from [1, 2a, 2b, 3, 4]. There are some other options that you can choose at the bottom of eval.py.

Directory structure

./agents: You can implement your own agent in this directory.
./benchmark: Data files for all the 4 tiers of the dataset

Adding your own agent

All you need to do is create an agent class with the method interact() or batch_interact().

​