README

README

Code to reproduce the results for the ICLR 2023 paper "Incompatibility
Clustering as a Defense Against Backdoor Poisoning Attacks"

https://arxiv.org/abs/2105.03692

To get started, first install requirements
> pip install -r requirements.txt
We used Python 3.8.1 for all experiments

Next, generate the poisoned datasets
> mkdir -p datasets
> python -m data.cifar10_backdoor
By default, this will generate the poisoned datasets used in the first set of
runs. The seeds used to generate the second and third sets of runs are also
provided in data/cifar10_backdoor.py. Our results are reported as the median of
the 3 runs.

To run our defense on a scenario
> python -m evaluation.backdoor_tests 0 resnet32
> python -m evaluation.backdoor_tests 23 preactresnet18
The required (positional) arguments are
  scenario_id   each poisoned dataset contains 24 scenarios, numbered 0-23
  architecture  either resnet32 or preactresnet18