FedPII

Official implementation of "Simple Yet Effective: Extracting Private Data Across Clients in Federated Fine-Tuning of Large Language Models"

🚀 Getting Started

1. Prepare Dataset

First, download the pre-processed .parquet dataset (available on Hugging Face) and place it in the data/processed/ directory.

Then, execute the following command to verify the dataset statistics:

python scripts/check_stats.py +data=sources

You should see a comprehensive dataset statistics report like this:

========================================
DATASET STATISTICS: fed_pii_v1.parquet
========================================
[Total Samples]: 16194
split
train 14575
test 1619
Name: count, dtype: int64
[Samples with PII]: 11633
[Breakdown by Task & Split]
count with_pii pii_ratio
split task test exam 224 77 34.38%
jud_read_compre 350 350 100.00%
jud_sum 265 265 100.00%
leg_case_cls 403 121 30.02%
sim_case_match 377 377 100.00%
train exam 2020 623 30.84%
jud_read_compre 3148 3143 99.84%
jud_sum 2386 2385 99.96%
leg_case_cls 3625 898 24.77%
sim_case_match 3396 3394 99.94%
========================================

Note on Raw Data Processing: If you are interested in how the raw data was processed into the final .parquet format, you can refer to scripts/make_data.py. Please note that the original raw data is not publicly provided due to privacy and licensing constraints. This script is included strictly for reference and transparency.

2. Federated Learning (FL) Partition

To simulate a federated learning environment, we partition the dataset across multiple simulated clients. Note that our partitioning algorithms are adapted from the open-source implementations in the FedLegal/FedLab repositories.

Prerequisite: The partitioning process relies on an encoding-clustering algorithm. You must download the xlm-roberta-longformer-base-16384 model in advance to compute the semantic embeddings.

Before proceeding, create a custom local configuration file under configs/local/ (e.g., private.yaml) to specify your local model paths:

# @package pathsmodel_root_external: "/path/to/your/models"

Once configured, run the following two-step process:

# Step 1: Unpack the datasets into an interim directory across tasks
python scripts/data_preparation.py
# Step 2: Generate Non-IID partitioned sample indices based on semantic clustering
python scripts/partition_data.py +partitioner=clustering +local=private

---

3. 🚧 Notice: Training & Attack Pipeline

The codebase for Federated Supervised Fine-Tuning, Evaluation, and PII Extraction Attacks is currently undergoing a privacy review and architectural refactoring. The remaining code will be released once this process is complete.

Citation

@inproceedings{hu-etal-2025-simple,
title = "Simple Yet Effective: Extracting Private Data Across Clients in Federated Fine-Tuning of Large Language Models",
author = "Hu, Yingqi and
Zhang, Zhuo and
Zhang, Jingyuan and
Wang, Jinghua and
Wang, Qifan and
Qu, Lizhen and
Xu, Zenglin",
booktitle = "Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics",
month = dec,
year = "2025",
address = "Mumbai, India",
publisher = "The Asian Federation of Natural Language Processing and The Association for Computational Linguistics",
url = "https://aclanthology.org/2025.findings-ijcnlp.113/",
pages = "1808--1827",
ISBN = "979-8-89176-303-6",
}

License

Shield:

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.