Applying a meta-learning mechanism based on cross-layer recognition in the Internet of Things network to promote federated learning - دانشکده فنی و مهندسی
Applying a meta-learning mechanism based on cross-layer recognition in the Internet of Things network to promote federated learning
نوع: Type: Thesis
مقطع: Segment: PHD
عنوان: Title: Applying a meta-learning mechanism based on cross-layer recognition in the Internet of Things network to promote federated learning
ارائه دهنده: Provider: Fazeleh Tavassolian
اساتید راهنما: Supervisors: Dr.Mahdi Abbasi, Dr.Abbas Ramezani
اساتید مشاور: Advisory Professors: Dr.Amir Taherkordi, Dr.Mohammad Reza Khosravi
اساتید ممتحن یا داور: Examining professors or referees: Dr. Muharram Mansoorizadeh, Dr. Ali Mohammad Latif , Dr. Parham Moradi
زمان و تاریخ ارائه: Time and date of presentation: 2025
مکان ارائه: Place of presentation: seminar room
چکیده: Abstract: In traditional Internet of Things (IoT) systems, the data stored on edge devices are typically transferred to a centralized server for training or updating machine learning models. However, accessing the necessary data for training accurate models can be challenging and costly, especially in cases where labeled data is limited. These challenges lead to two main issues: privacy concerns and the scarcity of sufficient data for training models in the field of machine learning. In this thesis, we present several federated machine learning models based on few-shot learning, including ResFed, to address these challenges, with a focus on improving model accuracy with limited data. In the proposed multi-stage federated meta-learning approach, edge nodes are trained within a centralized architecture, which leads to faster learning and model adaptation in scenarios with small datasets. In the proposed method, to mitigate the impact of limited data in the training dataset, data augmentation techniques are first used to increase the number of training samples. A feature extractor is then employed to learn a representation on the base set, and subsequently, classification is performed in the meta-learning stage. However, federated learning faces challenges such as bandwidth limitations and system and data heterogeneity across edge devices, which makes simultaneous model updates across all devices impractical. In this thesis, new methods for the optimal management of device resources and their effective selection are also introduced. The proposed approaches, including AdaptFFSL-DS and AdaptLightFedDS, leverage dual reinforcement learning to optimize device selection and adjust their participation levels in the learning process. These methods strike a balance between model accuracy and communication cost, taking into account system and data heterogeneity. Ultimately, the proposed AdaptLightFFSL-DSmethod, based on reinforcement learning, is introduced for optimal device selection in few-shot learning within a federated environment. This approach aims to improve the performance of learning models. The results from the implementation of these methods demonstrate that the proposed approaches can effectively address the challenges in federated learning with limited data and enhance system efficiency.
فایل: ّFile: Download فایل