Project Fiesole

Unique Image Identifiers: A proposal includes assigning unique numeric IDs to images and their respective quadrants. The hypothesis is that this would help the neural network learn patterns and relationships between various quadrants.
Training Variations: Different training methods using these IDs have been proposed, ranging from single-example training to batching. However, there are concerns about efficiency and computational speed.
Introducing Noise: I proposed the addition of random noise to the blacked-out sections of the image. This might simulate knowledge across the quadrants and improve the model&#x27;s accuracy.

Warning

This post is more than a year old. Information may be outdated.

TL;DR:

Can we VFL MNIST?

anaclumos/fiesole: Project Fiesole aims to demonstrate the feasibility of integrating CNN architectures within Vertical Federated Learning.

Project Fiesole aims to demonstrate the feasibility of integrating CNN architectures within Vertical Federated Learning to enhance data-driven models without compromising data privacy. In this context, data exists in islands, meaning various entities or organizations possess distinct features for the same sample population. Rather than consolidating these diverse datasets, Vertical Federated Learning allows for collaborative model training where each entity contributes its unique feature set. This approach not only upholds data privacy but also leverages the richness of diverse input parameters, offering a holistic perspective that a singular dataset might miss. Fiesole is a subproject of Project Florence.

Method

In VFL, each island has different "input parameters" for the same "subject". Thus, we split the original image (here, 9) into four distinct pieces in this demo. Each of them will necessarily work as a "distinct input."

63E892