Using Blockchain Technologies to Improve Security in Federated Learning Systems
Andrew R. Short, Helen C. Leligou, Michael Papoutsidakis, Efstathios Theocharis
Abstract
The potential of Federated Learning (FL) deployment increases rapidly as the number of connected devices increases, the value of artificial intelligence is recognized and networking technologies and edge computing evolves. However, as in any distributed system, a set of security issues arise in FL systems. In this paper, we discuss the use of blockchain technology to address diverse security aspects of FL systems and focus on the model poisoning attack for which we propose a novel Blockchain-based defense scheme. An assessment using data from the MNIST database has shown that the proposed approach, which has been designed to be implemented on blockchain technology, offers significant protection against adversaries attempting model poisoning attacks. The approach adopts a novel algorithm for evaluating the model updates, by verifying each model update separately against a verification dataset, without requiring information about the training dataset size, which is often unavailable or easily falsified.