Taeyoon Kim

Advancements in technologies such as massive Internet of Things (IoT) have led to the rapid proliferation of edge devices, generating vast amounts of diverse personalized local data that require large-scale distributed training of highly expressive but complex models to effectively capture varied data patterns. Split-federated learning (SFL) provides a basic framework for resource-constrained IoT devices to participate in federated learning (federated learning) with more expressive models by distributing computational load between clients and servers while preserving data privacy. However, existing SFL approaches still face scalability limitations: all participating clients handle only their front-end model segments while leaving the server to handle all remaining computations throughout the training process, intensifying server-side computational and memory bottlenecks, which in turn degrade training efficiency as the number of participating clients and model complexity increase. To address these issues, we propose parallel split-federated learning (Par-SF), which employs multipoint partitioning to distribute model blocks across idle clients, enabling group-based collaborative training that mitigates server-side bottlenecks by leveraging underutilized client resources. Additionally, we develop the client allocation scheme (CAS), a multiagent deep Q-network (MADQN)-based scheduler that optimizes group-based training by considering computational resources, network bandwidth, data volume, data freshness, and load balancing to maximize training efficiency. Extensive simulations across various training scales, model complexities, and data heterogeneity levels demonstrated that Par-SF significantly outperforms existing SFL approaches, achieving up to 2.82 times faster training speed while maintaining superior accuracy and stability, and reducing server computational and memory overhead by approximately 50 times, thereby ensuring scalability for large-scale IoT deployments with complex model training requirements.

Federated Learning (FL) presents a promising approach for collaborative intrusion detection while preserving data privacy. However, current FL frameworks face challenges with non-independent and identically distributed (non-IID) data and class imbalances in network security contexts. This paper introduces Dynamic Sampling-FedIDS (DS-FedIDS), a novel framework that enhances federated intrusion detection through adaptive sampling and personalization. DS-FedIDS extends the Federated Learning with Personalization Layers (FedPer) architecture by incorporating dynamic up/down sampling to address class imbalance issues in network security datasets. The framework maintains a global model for shared attack detection while enabling client-specific adaptation through personalized layers. Our approach effectively handles heterogeneous network environments, including Content Delivery Networks, Internet of Things, and industrial systems, each with distinct traffic patterns and attack profiles. Experimental results demonstrate that DS-FedIDS outperforms baseline FedPer in accuracy and efficiency, achieving superior detection rates across diverse attack classes while maintaining reasonable training overhead. Notably, DS-FedIDS excels in detecting minority attack classes and adapting to client-specific normal traffic patterns, making it ideal for real-world intrusion detection scenarios with inherently imbalanced and heterogeneous data distributions.

Federated learning (FL), in contrast to traditional centralized learning, has gained significant attention as it enables the training of high-performance neural networks by maintaining user data locally while exchanging model updates between the server and clients, i.e., end-to-end network exchange. However, in practical environment, when considering various anonymous participants’ computational resources, only a minority of clients can comply with the constraints imposed during the server-side training process. To address this issue, we propose Adaptive Scaling-Federated Learning (AS-Fed) based on Deep Q-Network (DQN) to dynamically orchestrate client’s local data, allowing the inclusion of a larger number of clients in the training procedure. Our experimental results demonstrate that the proposed AS-Fed approach outperforms the legacy scheme, achieving higher normalization performance during the training process.

The High Altitude Platform Station (HAPS) provides communication services within the stratosphere at an altitude ranging between 20 - 50 km. Utilizing the stable atmospheric conditions of the stratosphere, HAPS offers communication services over extensive areas while facilitating effective communication in line-of-sight (LoS) environments. This paper focuses on the analysis of HAPS cell coverage in a 3GPP standard based cellular communication network where multiple HAPS served as base stations based on the signal-to-interference-plus-noise ratio (SINR). Specifically, the study concentrates on whether user equipment (UE) positioned at the cell edge of the central HAPS maintains a downlink (DL) SINR above a certain threshold. Simulation results show that multiple HAPS 7-cell scenario coverage is achieved at 400 km ² and multiple HAPS one-cell scenario coverage is achieved at 230 km ² .

Federated learning (FL) is a promising collaborative learning approach in edge computing, reducing communication costs and addressing the data privacy concerns of traditional cloud-based training. Owing to this, diverse studies have been conducted to distribute FL into industry. However, there still remain the practical issues of FL to be solved (e.g., handling non-IID data and stragglers) for an actual implementation of FL. To address these issues, in this paper, we propose a cluster-driven adaptive training approach (CATA-Fed) to enhance the performance of FL training in a practical environment. CATA-Fed employs adaptive training during the local model updates to enhance the efficiency of training, reducing the waste of time and resources due to the presence of the stragglers and also provides a straggler mitigating scheme, which can reduce the workload of straggling clients. In addition to this, CATA-Fed clusters the clients considering the data size and selects the training participants within a cluster to reduce the magnitude differences of local gradients collected in the global model update under a statistical heterogeneous condition (e.g., non-IID data). During this client selection process, a proportional fair scheduling is employed for securing the data diversity as well as balancing the load of clients. We conduct extensive experiments using three benchmark datasets (MNIST, Fashion-MNIST, and CIFAR-10), and the results show that CATA-Fed outperforms the previous FL schemes (FedAVG, FedProx, and TiFL) with regard to the training speed and test accuracy under the diverse FL conditions.