[ Back ]


Resource-Aware Distributed Machine Learning for the Internet of Things

Author Eric SAMIKWA
Director of thesis Prof. Torsten Braun
Co-director of thesis
Summary of thesis

Internet of Things (IoT) systems generate large volumes of data from user devices. The raw data generated by IoT devices is very often private or sensitive and can be too large to transmit over the networks. For example, wearable devices such as Google Glass or Apple watch and medical sensors gather sensitive data by recording the daily activities of a user. This data is essential for executing Machine Learning (ML) models in order to deliver personalized services, early warning, and other intelligent IoT applications. Running ML models entirely on the cloud provides higher scalability because more resources for complex computation are readily available, but this approach shows three main disadvantages: Firstly, the time needed to send, process, and retrieve data from geographically distant data centers may not satisfy the real-time requirements of latency-critical applications. Secondly, the processing of raw sensor data on the cloud may expose sensitive information during data transmission, remote processing, and storage. Thirdly, transferring the raw sensor data from the IoT device to the cloud increases the ingress throughput on the backhaul network. Because of their proximity to the data, conventional consumer-level de- vices, such as IoT devices, are great candidates for the in-the-edge processing of the ML model. However, current state-of-the-art models such as Deep Neural Networks (DNN) have significant demands on memory, computation, and energy. This is incompatible with the resource-constrained nature of IoT devices. This research aims to determine how to efficiently distribute ML tasks across different elements in IoT systems, taking into account computation and communication constraints. Our work focuses on new resource-aware federated and split learning approaches that accelerates model training and inference time and minimizes energy consumption in resource-constrained hetero-generous IoT devices while taking into account the dynamic resources.

Status middle
Administrative delay for the defence 2024