↓Skip to main content

Explaining Black-Boxes in Federated Learning

·1 min

TL;DR #

Explaining Black-Boxes in Federated Learning adapts feature-attribution explanations to models trained collaboratively across multiple clients.

Problem #

Federated learning protects data locality, but this makes it difficult to apply centralized explainability methods. A server cannot simply collect all client data to explain the global model, and explanations produced independently by different clients may not be directly comparable.

What did we do? #

We designed a method that computes local explanations at the clients and combines the resulting feature attributions to explain predictions from the federated model. The approach builds on SHAP-style explanations while keeping the data on the clients.

Main findings #

A federated model can be explained without centralizing the records used to train it. Aggregating local attribution information provides a global view while preserving the distributed nature of the learning process.

Resources #

Cite this paper #

@InProceedings{10.1007/978-3-031-44067-0_8,
author="Corbucci, Luca
and Guidotti, Riccardo
and Monreale, Anna",
editor="Longo, Luca",
title="Explaining Black-Boxes in Federated Learning",
booktitle="Explainable Artificial Intelligence",
year="2023",
publisher="Springer Nature Switzerland",
address="Cham",
pages="151--163",
isbn="978-3-031-44067-0"
}