Towards Intelligent Participant Selection for Federated Learning in the Internet of Vehicles

<p dir="ltr">Federated Learning (FL) has gained considerable attention over the past decade or so for addressing key limitations in conventional centralized machine learning, i.e., particularly pertinent to data privacy, communication overhead, and latency. By decentralizing the training process, FL allows the clients (participants) to train the models locally and the clients, in turn, share only the model updates instead of sending the raw data to a central server. This approach, therefore, enhances data privacy and reduces communication costs, thereby making FL particularly valuable in privacy-sensitive domains, i.e., the Internet of Things (IoT), the Internet of Vehicles (IoV), healthcare, and finance. The decentralized approach underlying the FL paradigm facilitates IoV to enable secure and collaborative training not only amongst the vehicles but also between vehicles and pedestrians and roadside infrastructure while preserving privacy and improving response times which are crucial factors for time-sensitive applications, i.e., traffic management and collision avoidance. However, FL faces substantial challenges when applied to IoV, particularly, in the participant selection process. In case of IoV, wherein vehicles are the key participants, i.e., characterized by rapid mobility, dynamic environments, and high data and device heterogeneity, selecting the right participants is highly indispensable for optimizing the overall model performance. </p><p dir="ltr">This research thesis, therefore, begins with a comprehensive yet concise review of the state-of-the-art pertinent to FL in a bid to explore its core architectural aspects, i.e, scale of federation, data distribution, network topology, exchanged information, training procedures, machine learning models, and aggregation techniques. It also compares leading FL frameworks and datasets used for simulation and evaluation purposes. Following this, the thesis delineates how FL can empower various applications to reach their respective potential. The challenges pertinent to FL are then explored under two main categories, general challenges and those specific to the participant selection process, with a detailed discussion of the state-of-the-art participant selection processes that have been employed to address the later challenges. </p><p dir="ltr">The core contribution of this research thesis is envisaging of an intelligent framework entitled, FedCLF — Federated Learning with Calibrated Loss and Feedback Control, to address the significant challenges of the FL participant selection process in the context of an IoV network. The envisaged approach (a) enhances the overall model accuracy in case of highly heterogeneous data via the introduction of a new utility, i.e., calibrated loss, and (b) optimizes the resource utilization for resource-constrained IoV networks owing to the reduced sampling frequency of clients via a feedback control mechanism, thereby leading to increased efficiency in the FL process. FedCLF is evaluated vis-à-vis baseline models, i.e., FedAvg, Newt, and Oort, using CIFAR-10 dataset with varying data heterogeneity. The results depict that FedCLF significantly outperforms the baseline models by up to a <i>16% </i>improvement in high data heterogeneity-related scenarios with improved efficiency and resource utilization via reduced sampling frequency. </p><p dir="ltr">In conclusion, this research thesis not only advances the understanding of FL but also provides a practical solution in the form of a FedCLF framework which has been specifically tailored for IoV networks to tackle key challenges, i.e., highly dynamic environment, higher data heterogeneity, and resource limitations, thereby paving way for an effective, efficient, and scalable FL deployments in complex and heterogeneous environments.</p>