Protective optimization techonologies: case studies

This is the accompanying code to the paper "POTS: Protective Optimization Technologies".

It contains code for the two case studies in the latest version of the paper:

• Poisoning a credit-scoring model.
• Changine speed limits in a town to avoid Waze routing through it.

One POT in this repo that did not make into the paper:

• Evasion/"gaming" of the credit-scoring model.

Installation

System packages

On a Debian-based system these packages should have you covered:

apt install python3 python3-matplotlib python3-numpy python3-scipy python3-sklearn

Python

You need to have Python 3.6 or later. To install the packages, run:

pip install -r requirements.txt

Structure

• scripts --- a script for running the poisoning credit-scoring experiments, and the traffic routing POT.
• notebooks --- Jupyter notebooks for evasion and poisoning credit scoring experiments
• src --- common utilities and tools.
• images --- after running, the notebooks and scripts save plots here
• out --- the scripts saves simulation data here
• data --- German credit risk dataset

Running the poisoning experiments

PYTHONPATH=. python scripts/credit_poisoning.py

The experiments for evasion run fast, hence they are directly in the corresponding notebook.

Running the traffic routing experiments

Check the instructions here:

PYTHONPATH=. python scripts/anti_waze.py --help

Example run:

PYTHONPATH=. python scripts/anti_waze.py --town leonia run-one-experiment --target_time 5

Running notebooks

jupyter notebook

And choose the notebooks in the notebook folder

Citing

@inproceedings{KulynychOTG20,
  author    = {Bogdan Kulynych and
               Rebekah Overdorf and
               Carmela Troncoso and
               Seda F. G{\"{u}}rses},
  title     = {POTs: Protective Optimization Technologies},
  booktitle = {FAT* '20: Conference on Fairness, Accountability, and Transparency,
               Barcelona, Spain, January 27-30, 2020},
  pages     = {177--188},
  year      = {2020},
  url       = {https://doi.org/10.1145/3351095.3372853},
  doi       = {10.1145/3351095.3372853},
}