FlowChain: The Playground for Federated Learning in Industrial Internet of Things Environments

The Industrial Internet of Things (IIoI) lays the foundation for a new industrial revolution, the so-called Industry 4.0, in which every element, from machines to processes, is interconnected and fully automated, producing a huge amount of valuable data, crucial for making industrial processes more efficient and profitable. For this reason, it is common to run machine learning algorithms in order to extract useful information from these data. At the same time, due to bandwidth and privacy issues, it is often infeasible to transfer all these large-scale data to a centralized location where these algorithms can be performed. To address these scenarios, federated learning (FL) has been gaining ground in recent years. FL leaves the training data on the devices where they are produced and builds a global model by aggregating locally computed models, preserving user privacy and avoiding overwhelming the network with unnecessary raw data. However, there are still important challenges and limitations to the application of FL in Industry 4.0, mainly due to security issues and the fact that many solutions still suffer from single points of failure and bottlenecks. In this article, we present FlowChain, a framework that integrates FL with blockchain technology and decentralized identifiers (DIDs) to create an infrastructure that offers the possibility to easily exploit FL in Industry 4.0 scenarios. By using smart contract technology to automate the aggregation of partial models, it is possible to make FL fully decentralized. In addition, the fact that the blockchain is immutable and every transaction is verified and traceable makes FL secure. Furthermore, we exploit DIDs to be able to uniquely identify each element participating in industrial processes. This allows finer-grained control of authorizations so that only IIoT devices with a known and authorized identity can actually participate in the FL training process. Finally, we present some preliminary results that demonstrate the feasibility of the proposed approach.