Spring Conference
Authors Seyha Ros, Geonho Cha, Prohim Tam, Seokhoon Kim
Abstract
In wireless communications, advancements in end-device sensing capability drive a number of intelligent applications, particularly in real-time industrial imaging and visual-assisted services. Edge intelligence merges mobile edge computing (MEC) platforms and artificial intelligence (AI) modeling solutions to enable distributed resource-efficient model training and inference in 5G and beyond, allowing for real-time expectations and the application of AI-based models. However, maintaining the privacy of local data sharing and transmitting large amounts of raw data for creating edge models remains a challenge. As a result, federated learning (FL) is used in this research to provide a collaborative platform for building a final learning model with real-time image processing efficacy between the end device and the aggregator. Adaptive virtual aggregator resource placements strive to prioritize high mission-critical serving resources and reduce insufficient model dropouts due to the diversity of service types. For managing the time series analysis of local model parameter loading, a modified long short-term memory for congestion prediction (MLSTM-CP) is designed to weigh the configuration of resource pooling attributes in certain intervals. The installation of long-term forwarding paths of participant-aggregator selection and schedule-federated averaging of multi-service industrial imaging application classes is orchestrated using a trustworthy model aggregation policy. To compare the proposed model update solution to existing conventional policies, simulations are conducted.
0 Comments