Differentially Private Relational Learning with Entity-level Privacy Guarantees

This repository provides tools for fine-tuning large language models over sensitive graph-structured data for relational learning with entity-level differential privacy. It supports privacy accounting for coupled sampling in relational learning through node_dp_accounting module and prediction tasks on text-attributed graphs with privacy guarantees through private fine-tuning module.

Repository Structure

• node_dp_accounting module
  • pos_neg_amplification.py: Main script for running the RDP accountant on positive-negative sampling with adaptive gradient clipping.
• private fine-tuning module
  • Lora_SeqLP_prv.py: Main script for private fine-tuning with LoRA and relation prediction tasks.
  • arguments.py: Configurable arguments and CLI interface.
  • dataset.py: Dataset preprocessing and loading utilities.
  • model.py: Model wrapper with support for HuggingFace transformers.
  • trainer.py: Training loop with privacy tracking and evaluation.
  • transformers_support.py: Patches to adapt HuggingFace modules for compatibility with Opacus.
  • utils.py: Miscellaneous utility functions (e.g., logging, checkpointing).

Environment Setup

conda create -n pyvacy python=3.11
conda activate pyvacy
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118
pip install transformers==4.39.3
pip install sentencepiece
pip install peft
pip install datasets
pip install evaluate
pip install opacus
pip install wandb
pip install pandas
pip install scikit-learn

[Optional] Use with Jupyter Notebook

conda install ipykernel
ipython kernel install --user --name=pyvacy

Training Commands

To launch training, run:

bash lp_train_llama.sh <CUDA_ID> <DATASET_NAME> <EPSILON> <NOISE_SCALE> <CLIP_NORM> <BATCH_SIZE> <MODEL_NAME>

• MODEL_NAME options:
  • base -> bert-base-uncased
  • large -> bert-large-uncased
  • default -> meta-llama/Llama-2-7b-hf

The script wraps calls to Lora_SeqLP_prv.py using parameters defined in arguments.py.

Reproducing Table 1 Results

cd private_finetuning/script/
bash run_reproduce.sh

Notes

• This repository integrates Opacus with efficient per-loss-term gradient clipping and privacy accounting, compatible with HuggingFace’s transformer models.
• The noise scale used in training is computed based on a target privacy budget (ε), using the scripts provided in the node_dp_accounting/ directory.
• Make sure your wandb account is configured properly if logging is enabled.

Citation

Please cite our paper if you are interested in our work on differentially private relational learning.

Node/Entity-level DP
@article{huang2025differentially,
  title={Differentially Private Relational Learning with Entity-level Privacy Guarantees},
  author={Huang, Yinan and Yin, Haoteng and Chien, Eli and Wei, Rongzhe and Li, Pan},
  journal={arXiv preprint arXiv:2506.08347},
  year={2025}
}

Edge/Relation-level DP
@inproceedings{
    yin2024privately,
    title={Privately Learning from Graphs with Applications in Fine-tuning Large Language Models},
    author={Haoteng Yin and Rongzhe Wei and Eli Chien and Pan Li},
    booktitle={Statistical Foundations of LLMs and Foundation Models (NeurIPS 2024 Workshop)},
    year={2024},
}