A decentralized data evaluation framework in federated learning

Abstract

Federated Learning (FL) is a type of distributed deep learning framework in which multiple devices train a local model using local data, and the gradients of the local model are then sent to a central server that aggregates them to create a global model. This type of framework is ideal where data privacy is of utmost importance because the data never leave the local device. However, a major concern in FL is ensuring the data quality of local training data. Since there is no control over the local training data, ensuring that the local model is trained on clean data becomes challenging. A model trained on poor-quality data can have a significant impact on its accuracy. In this paper, we propose a decentralized approach using blockchain to ensure local model data quality. We use miners to validate each local model by checking its accuracy against a secret testing dataset. This is done using a smart contract that the miners invoke during the mining process. The local model is aggregated with the global model only if it passes a preset accuracy threshold. We test our proposed method on two datasets: the Brain Tumor Classification dataset from Kaggle, comprised of 7000 MRI images divided into two classes (Tumor/No Tumor), and the Medical MNIST dataset, which includes 58,954 images classified into six different classes: AbdomenCT, BreastMRI, ChestCT, Chest X-ray, Hand X-ray, and HeadCT. Our results show that our method outperforms the original FL approach in all experiments.