ginevracoal / adversarialGAN

Code for paper "Adversarial Learning of Robust and Safe Controllers for Cyber-Physical Systems"

Adversarial Learning of Robust and Safe Controllers for Cyber-Physical Systems

Code for paper "Adversarial Learning of Robust and Safe Controllers for Cyber-Physical Systems", Luca Bortolussi, Francesca Cairoli, Ginevra Carbone, Francesco Franchina, Enrico Regolin, 2020.

Abstract

We introduce a novel learning-based approach to synthesize safe and robust con- trollers for autonomous Cyber-Physical Systems and, at the same time, to generate challenging tests. This procedure combines formal methods for model verification with Generative Adversarial Networks. The method learns two Neural Networks: the first one aims at generating troubling scenarios for the controller, while the second one aims at enforcing the safety constraints. We test the proposed method on a variety of case studies.

Experiments

Code structure

• architecture/ contains the general GAN architecture and training / testing procedures
• utils/diffquantitative.py provides the logic to write, parse and check STL requirements
• utils/misc.py groups some minor helper functions
• model/ contains specific models for the different experimental setups, i.e. the attacker, the defender and the differential equations for the evolution of the system
• settings/ contains the initial configuration for each case study
• train_*, tester_* and plotter_* scripts execute, store and plot the simulations

Quick start

Once the repository has been cloned, create a python3 virtual environment and install the specified requirements.

pip3 install virtualenv
virtualenv -p python3 venv
source venv/bin/activate
pip install -r requirements.txt
cd src/

Code runs with Python 3.7.4. on Ubuntu 18.10.

Reproduce experiments

Change model settings in src/settings/*.

python train_*.py 
python tester_*.py -r=N_SIMULATIONS
python plotter_*.py -r=N_SIMULATIONS

Models and plots are saved in experiments/.

Cartopole with target example

python train_cartpole_target.py 
python tester_cartpole_target.py -r=1000
python plotter_cartpole_target.py -r=1000

Licence

Creative Commons 4.0 CC-BY