This is the code and package repository for the ACL 2024 Findings paper: DP-MLM: Differentially Private Text Rewriting Using Masked Language Models

Setup

Installation

You can install the package directly using:

pip install dpmlm

Optionally, you can install from source. In this repository, you will find a requirements.txt file, which contains all necessary Python dependencies.

Resource Bootstrapping

Before running the mechanism, you need to download the necessary NLTK libraries:

from dpmlm.utils import setup_resources

setup_resources()

Usage of DP-MLM

The core logic resides in the DPMLM class. You can now initialize it with custom calibration bounds to ensure the DP privatization is tuned to your specific model (and bounding strategy).

from dpmlm import DPMLM
from dpmlm.utils import calculate_logit_bounds, cleanup

# 1. (Optional) Calibrate bounds for your specific model (e.g., RoBERTa)
bounds = calculate_logit_bounds("FacebookAI/roberta-base")

# 2. Instantiate the mechanism
M = DPMLM(MODEL="FacebookAI/roberta-base", calibration=bounds, bound_strategy=None)

# 3. Rewrite text
private_text = M.dpmlm_rewrite("Hello world, this is a private text.", epsilon=25)

# 4. Cleanup (optional, when finished)
cleanup(model_instances=[M])

If you want to set a bounding strategy for the clip bounds (beyond simple min/max selection), you can do so by passing a lambda function:

# strategy as used in the paper
strategy = lambda mean, std, low, high: (mean, mean + 4*std)
M = DPMLM(MODEL="FacebookAI/roberta-base", calibration=bounds, bound_strategy=strategy)

DP-MLM Batched Mode

For longer documents, the batched mode provides significant performance increases by parallelizing masked token predictions on the GPU.

To use batching, simply run:

M.dpmlm_rewrite_batch("Large document text...", epsilon=25, batch_size=16)

Depending on your setup, you may need to tweak the batch_size parameter for the most optimal performance gains.

Handling Direct and Indirect Identifiers

The DP-IPI extension of DP-MLM follows from research work (see below) that targets specifically the privatization of indirect personally identifiers in texts.

To activate this feature, you must instantiate DP-MLM accordingly:

M = DPMLM(MODEL="FacebookAI/roberta-base", calibration=bounds, bound_strategy=None, IPI=TRUE, PII=True)

PII mode uses Presidio to mask out personally identifiable information, thereby following strict redaction and bypassing DP-MLM.

In IPI mode, only tokens tagged as indirect personal identifiers are privatized. This is determined by a specified fine-tuned IPI detection model. We provide a default one, but you can also specify your own with IPI_model.

In both cases, you must also set flags when calling rewrite functions, e.g.:

private_text = M.dpmlm_rewrite("Hello world, this is a private text.", epsilon=25, IPI=True, PII=False)

Input Document Length

As of a release in 2025, DP-MLM no longer has the shortcoming of the 512 token context window (256 with concatentation), which was due to the limitations of MLM context windows.

Now, DP-MLM operates with a sliding window, where the maximum context is given, centered around the target word to be privatized. Thus, DP-MLM now works on arbitrarily long documents!

Usage of other evaluated models

There is one other included file for replication of the paper, which is easily importable and reusable:

• LLMDP.py: implementations of both DP-Paraphrase and DP-Prompt. Note that for DP-Prompt, you will need to download the corresponding LMs, i.e., from Hugging Face.

M = LLMDP.DPPrompt()

M.privatize("hello world", epsilon=100)

Important note

In order to use LLMDP.DPParaphrase, you must download the fine-tuned model directory. This can be found at the following link: Model

Citation

Please consider citing the original work that introduced DP-MLM. Thank you!

@inproceedings{meisenbacher-etal-2024-dp,
    title = "{DP}-{MLM}: Differentially Private Text Rewriting Using Masked Language Models",
    author = "Meisenbacher, Stephen  and
      Chevli, Maulik  and
      Vladika, Juraj  and
      Matthes, Florian",
    editor = "Ku, Lun-Wei  and
      Martins, Andre  and
      Srikumar, Vivek",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2024",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-acl.554/",
    doi = "10.18653/v1/2024.findings-acl.554",
    pages = "9314--9328"
}

If you use the DP-IPI variant, please cite:

Coming soon!