Scholarlink

<p>Beyond the fifth generation (B5G) paradigm, including the Internet of Things (IoT), is projected to establish unprecedented scalability and real-time responsiveness. However, the large-scale deployment of such systems is prone to security attacks, which grow due to their decentralised architecture, minimal latency, and substantial data exchange. While traditional intrusion detection systems (IDS) struggle to effectively safeguard user privacy and system integrity in dynamic and distributed environments, large language models (LLMs) have recently unveiled novel domains in cybersecurity due to their proficiency in capturing intricate sequential patterns. This paper proposes a novel method for detecting intrusions specifically designed for IoT networks that utilize B5G technology. The system employs a Hierarchical Federated Learning (HFL) model in conjunction with lightweight LLMs, such as TinyLLaMA and DistilBERT, to ensure that model training remains private, scalable, and efficient across diverse devices. In particular, the architecture consists of three tiers: clients, edge aggregators, and a central server, which, as a result facilitates hierarchical model aggregation while preserving data locality. The framework integrates quantisation, knowledge distillation, and Low-Rank Adaptation (LoRA) to improve the efficiency of edge deployment. Simulation results reveal that the proposed approach outperforms the conventional federated learning baselines on the TON IoT dataset. In addition, the proposed framework effectively accommodates non-IID data distributions and reduces communication overhead by 28%, indicating its potential functionality in real-world B5G IoT intrusion detection.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight large language model based hierarchical federated learning for B5G enabled IoT intrusion detection networks

  • 期刊论文

  • 发表日期：

    2026年3月4日

查看全文
  • Rami Mohawesh,
  • Haitham Al-Obiedollah,
  • Sumbal Maqsood,
  • Haythem Bany Salameh

摘要

Beyond the fifth generation (B5G) paradigm, including the Internet of Things (IoT), is projected to establish unprecedented scalability and real-time responsiveness. However, the large-scale deployment of such systems is prone to security attacks, which grow due to their decentralised architecture, minimal latency, and substantial data exchange. While traditional intrusion detection systems (IDS) struggle to effectively safeguard user privacy and system integrity in dynamic and distributed environments, large language models (LLMs) have recently unveiled novel domains in cybersecurity due to their proficiency in capturing intricate sequential patterns. This paper proposes a novel method for detecting intrusions specifically designed for IoT networks that utilize B5G technology. The system employs a Hierarchical Federated Learning (HFL) model in conjunction with lightweight LLMs, such as TinyLLaMA and DistilBERT, to ensure that model training remains private, scalable, and efficient across diverse devices. In particular, the architecture consists of three tiers: clients, edge aggregators, and a central server, which, as a result facilitates hierarchical model aggregation while preserving data locality. The framework integrates quantisation, knowledge distillation, and Low-Rank Adaptation (LoRA) to improve the efficiency of edge deployment. Simulation results reveal that the proposed approach outperforms the conventional federated learning baselines on the TON IoT dataset. In addition, the proposed framework effectively accommodates non-IID data distributions and reduces communication overhead by 28%, indicating its potential functionality in real-world B5G IoT intrusion detection.