FREE6G: Free-cell networks based on machine learning (redes cell-free basadas en aprendizaje automático)

Ref: CTTC-2022-20037

Overview

This project aims to design decentralized solutions based on blockchain technology to improve security and privacy in federated learning.

Context

Machine learning models are currently ubiquitous. These models have been infiltrating different facets of society to the point where they can be found in diverse areas such as economics (e.g. credit scoring systems), education (e.g. in the creation of virtual tutors that adjust their messages based on students’ performance in subjects), leisure (e.g. movie, series, or music recommendation systems), and the penal system (e.g. the use of COMPAS to make decisions about granting parole to prisoners), to name just a few paradigmatic examples.

Traditional machine learning models are built from a training dataset, which is usually collected prior to the learning process and is available in the system where the model is trained. That is, both the training dataset and the model creation process are completely centralized. This centralization leads to certain problems, such as a lack of scalability when creating large models or concerns about data privacy included in the training. The latter aspect is particularly critical when data is linked to individuals or can reveal industrial secrets, and motivates the creation of federated learning architectures, in which the training data is stored on end users’ devices.

Thus, federated learning systems are capable of learning a collective model based on data from different devices while keeping the data on individual devices.

Federated learning architectures can be centralized, with a server acting as a coordinator in the learning process, receiving model updates calculated by individual devices and combining them to create the collective model. However, there are also decentralized architectures, in which the coordinator is dispensed with, and collaborating devices coordinate with each other to create the model. It is in the context of this decentralized architecture that synergies with blockchain technology appear.

The integration of blockchain technology with federated learning models is still in its early stages. The aim of this work is to expand on these preliminary works by designing decentralized strategies based on blockchain for improving privacy and security in federated learning environments. While previous works have focused on the benefits of integration for other purposes (e.g., to incentivize devices to participate) or focused on specific scenarios (e.g., autonomous vehicles), this work aims to advance the improvements that the integration of both technologies can bring in terms of the security and privacy of FL environments.