This repository contains code related to the paper Machine Unlearning of Features and Labels published at NDSS 2023 and is structured as follows:

• We tested the code with python3.7.7
• We recommend setting up a virtual environment (e.g. using virtualenv)
• Install depdendencies via pip install -r requirements.txt
• Install optional dependencies (notebooks etc.) via pip install -r opt_requirements.txt
• For the backdoor experiments, have a look at example_notebooks/Cifar_data.iypnb. This notebook shows how to setup the data as expected by our backdoor experiments.

• The Code for the unleraning strategies is contained in the Unlearner folder. The DNNUnlearner class contains the first- and second order update strategies, all other classes inherit from it.
• The Applications folder contains some examples how to use the Unlearner classes as discussed in Section 6 of the paper.

Due to size limitations we publish not every model but some to experiment with.

• The LSTM folder contains two language generation model as described in the paper. The canary sentence has been inserted 8 and 29 times respectively and the telephone number that will be predicted is 0123456789.
• The CNN folder contains the poisoned CNN model that has been trained on the CIFAR-10 dataset.

We provide examples to reproduces the results from the paper in jupyter notebooks.

Due to size limitations we did not upload the raw data for the Drebin and Enron dataset and refer to the original papers instead. The vector representations to run the experiments are given instead. All of them can be found in the train_test_data folder.

If you found any of this helpful please cite our paper. You may use the following BibTex entry

@inproceedings{WarPirWreRie20,
    title={Machine Unlearning of Features and Labels},
    author={Alexander Warnecke and Lukas Pirch and Christian Wressnegger and Konrad Rieck},
    year={2023},
    booktitle={Proc. of the 30th Network and Distributed System Security (NDSS)}
  }