Data Poisoning Attacks on Regression Learning and Corresponding Defenses

This repository contains the code for the paper 'Data Poisoning Attacks on Regression Learning and Corresponding Defenses' (see ArXiv.org for full paper), published at PRDC2020. This repository implements the 'Flip' data poisoning attack on regression learning, and the 'Trim' and 'iTrim' defense. It evaluates these on 26 data sets.

How to run:

1. Download and include the required data sets (see section Datasets below).
2. Set the python path of main.py to the base of this directory.
3. Set settings.no_attack_run = True and run main.py. This will create the training data and split accordingly.
4. Set settings.no_attack_run = False and run main.py. The code will parallelize all experiments over the available combinations of data sets / regressors and finally output the results to the console.

Datasets:

Unfortunately, we're not able to prodive the data sets used in the repository (due to potential copyright issues). Thus, we list what data sets we used and where to obtain them:

In the folder res, place the following data sets:

• Github_regression: https://github.com/paobranco/Imbalanced-Regression-DataSets
• Keel Data sets: https://sci2s.ugr.es/keel/datasets.php
• Three data sets provided by [1]: Warfarin [2], loan and armesHousing.

The file res/tree.txt provides the layout of the folder.

Citation

If you use this work, please cite us:

Nicolas M. Müller, Daniel Kowatsch and Konstantin Böttinger. Data Poisoning Attacks on Regression Learning and Corresponding Defenses - IEEE Pacific Rim International Symposium on Dependable Computing (PRDC) 2020

Bugs

If you find bugs, feel free to fix them and send us a pull request.

References

[1] Jagielski, M., Oprea, A., Biggio, B., Liu, C., Nita-Rotaru, C., and Li, B. Manipulating Machine Learn-ing: Poisoning Attacks and Countermeasures for Regression Learning. InProceedings - IEEE Symposium on Security and Privacy, volume 2018-May,pp. 19–35. IEEE, may 2018. ISBN 9781538643525.doi:/10.1109/SP.2018.00057.URL/https://ieeexplore.ieee.org/document/8418594/.

[2] IWPC Data set: https://www.pharmgkb.org/downloads