Implementation of Opportunistic Federated Learning System for Heterogeneous Resource Environments

High-performance mobile devices equipped with various sensors capable of executing machine learning tasks have recently become widely available. Generally, sensing data is aggregated on a central server to train machine learning models. However, this approach consumes network and computational resources due to data aggregation and model training, while also raising privacy concerns related to data sharing. Alternatively, group construction methods for distributed machines, leveraging based on the location information, have been proposed to enable collaborative learning among mobile devices. However, tasks expected to run on mobile devices are diverse, and it is necessary to consider environments where various tasks coexist. Moreover, in real-world environments, devices vary in computational per-formance, communication performance, and the data they store, making it essential to address the impact of these factors on task execution. In this paper, we propose an opportunistic federated learning system that adaptively assigns groups for executing federated learning tasks based on the status of devices. We implement the proposed system and evaluate its performance, along with its cooperative device candidate selection algorithms.