Titulo Estágio
Low Power Architecture for Federated Learning in IoT Edge Image-based Processing
Áreas de especialidade
Sistemas Inteligentes
Engenharia de Software
Local do Estágio
Instituto de Telecomunicações (IT) e Instituto de Sistemas e Robótica (ISR)
Enquadramento
The recent proliferation of machine learning methods for automatic image classification, detection, segmentation, and scene understanding, provides an unprecedented opportunity for the rise of cloud-based real-time image processing algorithms with distributed IoT image acquisition nodes. In this work, we propose a novel distributed architecture for real-time intelligent image processing at the edge of the low power IoT network, leveraging the concept of federated learning, which decentralizes the training process of the model to safeguard users’ privacy and security.
For this, we will use dedicated and highly parallel hardware at the edge of the IoT network to perform training on real-time image data collected by distributed IoT acquisition nodes. We explicitly incorporate data preparation steps and perform dimension reduction to optimize bandwidth and computational cost. Moreover, a distribution of the trained models to client applications will be achieved via a cloud interface, providing for real-time prediction on unseen data.
To assess the performance of the novel architecture proposed, we will define a set of case study applications, such as medical imaging abnormality detection, real-time behavioral data pattern analysis of elderly people, or image segmentation to support forestry cleaning operations. Finally, a detailed comparison of results, accuracy, throughput, energy, and latency will be conducted.
Objetivo
• Propose a novel distributed architecture for real-time intelligent image processing at the edge of a low power IoT network.
• Explore the concept of federated learning, decentralizing the training process of the model to safeguard users’ privacy and security.
• Perform training on real-time image data collected by distributed IoT acquisition nodes at the edge of the IoT network using highly parallel hardware, performing dimension reduction to optimize bandwidth and computational cost.
• Distribute the models to client applications via a cloud interface, providing for real-time prediction on live images.
• Tests, adjustments, and validation of the proposed architecture with detailed assessment of results, accuracy, throughput, energy, and latency.
Plano de Trabalhos - Semestre 1
• Literature review on existing relevant systems, specially focused on Federated Learning.
• Familiarization and development of the basic IoT image acquisition nodes and the highly parallel hardware at the network edge, and their evaluation concerning the adequacy and convenience for the proposed tasks.
• Implementation of secure and scalable real-time data acquisition and communication of (possibly low resolution) images to the highly parallel hardware at the edge of the IoT network, encompassing bandwidth optimization measures.
• Preliminary training models.
• Write the Intermediate Report.
Plano de Trabalhos - Semestre 2
• Develop training methods for image classification/segmentation, according to the datasets used and case study defined, encompassing dimension reduction.
• Data integration and cloud management of the models, and connection to the client terminals for real-time prediction on live images.
• Iteratively test, refine, and validate the system, providing a stable version of the architecture. Thorough analysis of results and performance with real world images (either via datasets or real time acquisition). assessing failure rate of communications and basic functionalities, test the security and scalability of the system as well as detailed assessment of accuracy, throughput, energy, and latency.
• MSc. Dissertation writing and scientific paper on the subject.
Condições
This methodology will result in a Low Power Architecture for Federated Learning in IoT Edge Image-based Processing with proven performance in terms of results in a experimental prototype, which will be demonstrated by the student and document in his MSc. Dissertation.
Observações
This Dissertation work will take place at the Institute of Telecomunications (IT) ant the Institute of Systems and Robotics (ISR)
Main Supervisor (IT): Prof. Gabriel Falcão, gff@deec.uc.pt
Co-supervisors (ISR): Prof. Paulo Peixoto, peixoto@deec.uc, and Doctor David Portugal, david.portugal@deec.uc.pt
Orientador
Gabriel Falcão
gff@deec.uc.pt 📩