BANND

A project for the course "Trustworthy Machine Learning". We investigate defenses for backdoor attacks on neural networks during training.

Dependencies

The project requires Python 3.10, and dependencies are installed using pip.

To install the dependencies, run python -m pip install -r requirements.

We recommend using a Virtual Environment for the project. To do that, install virtualenv with python -m pip install virtualenv and create a new virtualenv in the project with python -m virtualenv ./venv. Then, activate the virtualenv with source ./venv/bin/activate and run all the python and pip commands within it.

Running

Run the project with python bannd.py. Pass the -h/--help flag to see all the parameters. Note that for brevity's sake, some parameters like the number of epochs and batch size are omitted and can't be passed.

We support 3 modes of running:

• Baseline: pass --runtype baseline to see the network's baseline accuracy without any attack or defense.
• Attack: pass --runtype attack and --poison_rate Y (any other number between 0.01 and 0.99, we tested with 0.1, 0.3, and 0.5) to run the badnets attack on the network, by poisoning some percentage of the samples with the backdoor. Expect the attack success rate to increase with time.
• Defense: pass --runtype defense, --poison_rate X, and --quantile-threshold Y (any number between 0.00 and 0.99, we tested with 0.0, 0.5, 0.7, and 0.85) to run the attack and defend against it.

Examples

To run the baseline:

python bannd.py --runtype baseline

To run the attack:

python bannd.py --runtype attack --poison_rate 0.1

To run the defense:

python bannd.py --runtype defense --poison_rate 0.1 --quantile-threshold 0.7