Edge model: An efficient method to identify and reduce the effectiveness of malicious clients in federated learning

Abstract

In the rapidly changing world of data-driven decision making, machine learning techniques have become increasingly popular. However, traditional machine learning models often rely on centralized data repositories, which can pose challenges to privacy and security. To address these limitations, federated learning has emerged as a groundbreaking approach that uses distributed data while preserving privacy. Federated learning, also known as collaborative machine learning, is a decentralized learning paradigm that allows multiple devices or entities to work together to build a shared model without sharing their raw data. Instead of sending sensitive data to a central server, federated learning enables devices to train models locally and share only aggregated updates to the model. This distributed approach empowers organizations to leverage the collective intelligence of their devices while ensuring the privacy of their data. in this paper, we have tried to present a method that is a new way to deal with malicious users and also reduce their effectiveness on the global model. The idea of this paper is a combination of using several global models instead of one global model and identifying clients in these models and removing them from the learning process. In this approach, certain global models are trained using learning algorithms and input from local users. Consequently, instead of having a single global model, we will have a collection of global models. When it comes to predicting the label of a test sample, the selected model will be determined based on the highest accuracy percentage achieved in label prediction among the available global models after eliminating any potentially harmful clients. We will show that the method presented in this paper will significantly reduce the impact of malicious users. We have evaluated this method on the popular MNIST, Cifar10 and Fashion MNIST datasets. The evaluation results show that the proposed method significantly reduces the effect of malicious users.