Privacy-aware secure explainable data-driven models in federated learning

Federated learning has received a lot of attention lately. It seems a good collaborative approach to build data-driven models. Agents or clients help themselves to build the model without sharing the data. Unfortunately, it has been shown that this type of framework does not avoid privacy and security risks.

Data-driven models in federated learning are usually simple in the sense that they are numerical models based on deep learning.

Data is distributed, but the type of data is similar in different agents (often either vertically or horizontally distributed). Explainability is limited or absent. In this project we will work on two key extensions not usually considered in the federated learning literature.

First, we will extend the current state-of-the-art considering agents that have data with completely different structure. That is, highly heterogeneous data.

Second, we will provide explainability either by means of building transparent models (beyond standard numerical deep learning models) or by means of explanations based on training data and data provenance. Solutions will be private and secure by design.

More concretely, we will develop privacy-aware secure explainable data-driven models within a federated learning scenario.

We plan to develop tools to support privacy preserving and secure data integration, and implement a framework to provide secure transparent data-driven models as well as provenance-rich explainability.