DistFL: An Enhanced FL Approach for Non Trusted Setting in Water Distribution Networks

The Internet of Things (IoT) is changing today's world, and Machine Learning (ML) is a major contributor to this revolution in terms of data exchange to mature connected objects. In this context, federated learning (FL) is emerging, a new ML paradigm that drives a model on decentralized data, which can be distributed across many IoT devices. FL has grown considerably in recent years, both in academia and industry. However, the majority of FL algorithms assume that all client nodes are honest and willing to participate in cooperative model learning. Thus, each node is able to provide reliable local models to the central server. However, in real-life scenarios, nodes may be corrupt, malicious, or both, and may not cooperate fairly during training phases. In this paper, we address the above challenge by proposing a new FL algorithm called DistFL. The main objective of DistFL is to prevent biased training by identifying malicious nodes during the training phase. We evaluate the effectiveness of our technique and demonstrate it through a concrete implementation, comparing DistFL with conventional FL. Even with up to 50% malicious nodes, the runtime cost of the DistFL model is still better than that of the conventional FL model, and its final accuracy reaches 97% with a loss function convergence rate twice that of the conventional FL model. For this study, we used urban water data to deal with leakage and distribution faults in the water network among different end users in this area.