Project data

Funding entity: Italian Ministry of University and Research

Call: PRN 2017

Coordinator: UNIVERSITA’ DEGLI STUDI DI CAGLIARI

UNISI Principal Investigator: Stefano Melacci

Department: Information engineering and mathematics

Start date: 29 August 2019 – End date: 29 August 2023

 

Description

The REXlearn project aims at improving Machine Learning technologies, focussing on three main challenges that are hindering current progress towards the development of more secure algorithms.

The first challenge is the challenge of security evaluation. The project aims at adapting techniques from security engineering and computer security to devise a proper threat model that allows envisioning
potential attacks against learning algorithms.

The second challenge posed in REXlearn is the challenge of countering attacks to Machine Learning-based models. The design of effective defenses demands for learning paradigms that enable incorporating models of potential attacks into the learning process and  seek stable and reliable equilibrium strategies.

The third challenge is the challenge of designing interpretable machines. The project aims at addressing this challenge by studying the properties of popular learning algorithms to identify the most influential features and prototypes that explain their local predictions and global behaviour. The aforementioned challenges are faced in the context of  different application scenarios, including malware detection and object recognition in images and videos.

The learning algorithms developed in this project can have a significant impact to tackle challenges that demand for stable and reliable learning paradigms, as those faced by recent computervision
and deep-learning technologies.

 

This project has received funding from Ministry of University and Research (MUR) – PRIN 2017