Master thesis 2020 - Data Science & Scientific Computing

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

Abstract

Data-driven models are bringing huge advantages in terms of performance and effectiveness in many sectors. In the field of Cyber-Physical Systems, although, they are not widespread since their opaqueness and lacks formal proofs pose a serious threat to their employment.

We propose and analyze a novel approach in creating robust and safe controllers drawing from the literature of the adversarial learning.

Our method trains in an adversarial way two Neural Networks to reach a twofold goal: obtain one network that is able to generate difficult configurations of the environment and another that is able to overcome them in a safe e robust way. The aim is to create a formally verified controller and, at the same time, to give insights on the most demanding corner cases of a given model.

The approach is promising and worthy of further investigation.

Project structure

The project has been developed in PyTorch and still needs to undergo a heavy refactoring. The main directories are:

doc, contains the Latex source code of the whole thesis
misc, contains the thesis and the presentation slides in PDF
src, the actual source code of the architecture developed

Source code

The project provides 3 main modules to create the desired experimental setup.

architecture.py provides the main structures of the whole concept (NNs, training and testing procedure)
diffquantitative.py provides the logic to write, parse and check STL formulae
misc.py groups some minor helper functions

Each experimental setup is composed of:

the model of the world that includes the definition of the attacker and defender and the differential equation that describes their evolution over time (model_*.py)
the training setup that defines the configuration of the architecture employed in the experiment (train_*.py)
(optional) the testing setup that evaluates the trained model in controlled scenarios and stores the experimental data (tester_*.py)
(optional) the plotting scripts that read the experimental data and produce the required visualizations (plotter_*.py)

Note: the tests performed by the class architecture.Tester can only measure the performance of the training process, they are not useful to assess the experimental setup.

Quick start

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

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Once the environment is ready, it's possible to launch the training of a specific model:

cd src
python train_platooning.py

Disclaimer

The code, despite its robustness, is far from being production-ready: it needs to be refactored and, in some parts, it requires a redesign to allow a faster prototypation of experimental models and setups. Therefore it should be considered as a Proof Of Concept.

Licence

Creative Commons 4.0 CC-BY

ffranchina / master_thesis_2020