Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/sap-samples/security-research-dp-hierarchical-text

SAP Security Research sample code to reproduce the research done in our paper “On the privacy-utility trade-off in differentially private hierarchical text classification".
https://github.com/sap-samples/security-research-dp-hierarchical-text

research sample sample-code security

Last synced: about 2 months ago
JSON representation

SAP Security Research sample code to reproduce the research done in our paper “On the privacy-utility trade-off in differentially private hierarchical text classification".

Awesome Lists containing this project

README 

        
# Differentially Private Hierarchical Text Classification



[![REUSE status](https://api.reuse.software/badge/github.com/SAP-samples/security-research-dp-hierarchical-text)](https://api.reuse.software/info/github.com/SAP-samples/security-research-dp-hierarchical-text)



## Description



SAP Security Research sample code to reproduce the research done in our paper "On the privacy-utility trade-off in

differentially private hierarchical text classification"[1].



## Requirements



- [Python](https://www.python.org/) 3.7

- [Tensorflow Privacy](https://github.com/tensorflow/privacy)

- [Tensorflow](https://github.com/tensorflow)

- [transformers](https://github.com/huggingface/transformers/)

- [numpy](https://numpy.org/)

- [scipy](https://scipy.org/)

- [pandas](https://pandas.pydata.org/)

- [anytree](https://anytree.readthedocs.io/)

- [scikit-learn](https://scikit-learn.org/)

- [matplotlib](https://matplotlib.org/)

- [tqdm](https://tqdm.github.io/)



## Download and Installation



### Differentially Private Hierarchical Text Classification Framework



Implementation of several hierarchical text classification (HTC) neural networks

and the corresponding differential privacy (DP) adversary to quantify information leakage in the trained HTC models.



### Install



Running `make install` in the root folder should be enough for most use cases.



The command will create the basic project directory structure and installs dph as well as other requirements.

You can use pip as your package manager and install the `dph` package via `python -m pip install -e ./`

For other package managers you need to install dpa using `setup.py`.



### Directory Structure



After having run `make install`, the following directory structure should be created in your local 

file system. Note: Everything that must not be tracked by git is already in `.gitignore`.



```

DPAttack/

     |-- Makefile

     |-- setup.py

     |-- requirements.txt

     |-- data/              # dataset files

     |-- logs/		        # experiment logs 

     |-- notebooks/         # evaluation notebooks

     |-- dph/			    # source root

          |--core/	            # hierarchical text classification framework

          |--mia/	            # membership inference attack framework

          |--projects/	        # project implementations using dph

          |--utils/             # utility modules



```



### Using the Framework



For every dph project (dataset), a subdirectory in `./dph/projects` is recommended.

Inside the folder, a subclass of an [`HTC Experiment`](dph/core/experiment.py) should be created as it is done for the existing datasets.

To train an HTC model, call the `start` method of an experiment.

To train a DP-HTC model, pass an instance of [`DPParameters`](dph/core/parameters/parameters.py) to the `Experiment` constructor.



To train an attack model for an (DP-)HTC model, a [`AttackExperiment`](dph/mia/mia_experiment.py) has to be created.

To start an attack, call the `start` method of the attack experiment.



### Reproducing the tables and diagrams of the paper



In the following, we explain how the different tables and diagrams of the paper can be reproduced. 

Running the python scripts for [BestBuy], [Reuters] and [DBPedia] will execute the different project's experiments.

The scripts' parameters can be adapted to execute all possible scenarios that we examined in the paper. 

The resulting experiment dictionary in the `logs` folder will contain several json files with the target/attack model metrics (train and test).

The results can then be retrieved and plotted in a notebook, as we have done in [`notebooks/plots-dp.ipynb`](notebooks/plots-dp.ipynb).



[BestBuy]: dph/projects/bestbuy/mia/experiment.py

[Reuters]: dph/projects/reuters/mia/experiment.py

[DBPedia]: dph/projects/dbpedia/mia/experiment.py



### Contributors



- Dominik Wunderlich

- Daniel Bernau



### Contributing



If you wish to contribute code, offer fixes or improvements, please send a pull request. Due to legal reasons, contributors will be asked to accept a DCO when they create the first pull request to this project. This happens in an automated fashion during the submission process. SAP uses [the standard DCO text of the Linux Foundation](https://developercertificate.org/).



If you want to report issues or have questions about the project, please open an issue in the GitHub repository or contact the maintainers directly via email. We will respond as soon as possible.



## Citations

If you use this code in your research, please cite:



```

@article{WBA+21,

      author    =   {Dominik Wunderlich and 

                    Daniel Bernau and 

                    Francesco Aldà and 

                    Javier Parra-Arnau and 

                    Thorsten Strufe},

      title     =   {On the privacy-utility trade-off in differentially private hierarchical text classification}, 

      eprint    =   {2103.02895},

      archivePrefix={arXiv},

      url       =   {http://arxiv.org/abs/2103.02913},

}

```



## References

[1] Dominik Wunderlich, Daniel Bernau, Francesco Aldà, Javier Parra-Arnau, Thorsten Strufe:

On the privacy-utility trade-off in differentially private hierarchical text classification.

arXiv:2103.02895

http://arxiv.org/abs/2103.02913



## Code of conduct



See the [Code of conduct tab](https://github.com/SAP-samples/security-research-dp-hierarchical-text?tab=coc-ov-file).



## License



Copyright (c) 2024 SAP SE or an SAP affiliate company. All rights reserved. This project is licensed under the Apache

Software License, version 2.0 except as noted otherwise in the [LICENSE](LICENSES/Apache-2.0.txt) file.