1-Diffractor

1-Diffractor is a high-performance library for word-level text perturbation leveraging Metric Differential Privacy. It maps text into 1D sorted embedding spaces to apply noise, ensuring privacy guarantees while maintaining semantic utility.

Key Features

• Metric DP Implementation: Support for both Truncated Geometric and Truncated Exponential (TEM) mechanisms.
• Automated Embedding Management: Automatically downloads and caches filtered embedding models (GloVe, Word2Vec, Numberbatch).
• Parallel Processing: Uses optimized multiprocessing to perturb large batches of text quickly.
• BYOE (Bring Your Own Embeddings): CLI tools to clean and integrate custom embedding files into 1-Diffractor.

Quickstart Guide

Installation

pip install dp-diffractor

Basic Usage

from diffractor import Diffractor, DiffractorConfig

# Configure the privacy mechanism
config = DiffractorConfig(
    method="geometric", 
    epsilon=1.0, 
    verbose=True
)

with Diffractor(config) as df:
    texts = ["Differential Privacy is really cool!", "Hello world."]
    perturbed = df.rewrite(texts)
    print(perturbed)

Optionally, you can also pass an epsilon directly to rewrite without having to instantiate the mechanism again. Additionally, you may pass a list of epsilon values, where the length of this list much match the token count of the text exactly, as per nltk.word_tokenize.

Note: due to the nature of the word embedding models, the input text must be all lowercase! We perform an extra check for this as well and convert the inputs to lowercase.

Advanced Configuration

The DiffractorConfig object allows you to customize the privatization parameters:

Managing Embeddings

1-Diffractor keeps a local cache (by default, ~/.cache/diffractor) to store embedding files.

Custom Embeddings (BYOE)

If you have your own embedding file, you must filter it against the internal vocabulary to ensure it works with the privatization mechanism:

# In your terminal
diffractor-clean path/to/my_vectors.txt

Then, use it during startup:

df = Diffractor(model_names=["my_vectors_filtered"])

Default Models

By default, 1-Diffractor fetches and uses the following embedding models:

• conceptnet-numberbatch-19-08-300
• glove-twitter-200
• glove-wiki-gigaword-300
• glove-commoncrawl-30
• word2vec-google-news-300

Citation

If you find 1-Diffractor useful or make use of it in your research, please be sure to cite the original paper:

@inproceedings{10.1145/3643651.3659896,
author = {Meisenbacher, Stephen and Chevli, Maulik and Matthes, Florian},
title = {1-Diffractor: Efficient and Utility-Preserving Text Obfuscation Leveraging Word-Level Metric Differential Privacy},
year = {2024},
isbn = {9798400705564},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3643651.3659896},
doi = {10.1145/3643651.3659896},
booktitle = {Proceedings of the 10th ACM International Workshop on Security and Privacy Analytics},
pages = {23–33},
numpages = {11},
keywords = {data privacy, differential privacy, natural language processing},
location = {Porto, Portugal},
series = {IWSPA '24}
}

Please also consider citing the hosted embedding files:

@dataset{meisenbacher_2026_19701515,
  author       = {Meisenbacher, Stephen},
  title        = {Filtered Embedding Files for 1-Diffractor},
  month        = apr,
  year         = 2026,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.19701515},
  url          = {https://doi.org/10.5281/zenodo.19701515},
}