DeepBern-Nets: Taming the Complexity of Certifying Neural Networks using Bernstein Polynomial Activations and Precise Bound Propagation
DeepBern-Nets is class of Neural Networks with Bernstein Polynomial based activation functions. Bernstein polynomials exhibit several key properties that allows tight computation of output bounds of all neurons in the network making them highly amenable to certification tasks. More details can be found in the paper DeepBern-Nets: Taming the Complexity of Certifying Neural Networks using Bernstein Polynomial Activations and Precise Bound Propagation.
We recommend using the provided conda environment file to install all dependencies. To create a conda environment, run
conda env create --name deepbern --file=environment.yml
conda activate deepbern
Make sure the environment was created without errors.
The main training file is train.py which accepts a YAML configuration file. For example, to run the code using one of the example configs, run
python train.py configs/mnist_FCNNb_clean.yml
There are 3 supported training modes:
- Certified training mode (using Bern-IBP bounding)
- Adversarial training mode (using PGD attacks)
- Clean training mode
The training mode can be specified by modifying the TRAIN.MODE parameter in the configuration file. To see examples of different training modes and their usage, refer to the sample configuration files located in the configs directory. The models are saved in the designated experiments folder, the location of which can be configured within the configuration file.