Scholarlink

<p>Anomaly detection plays a vital role in ensuring the safety, reliability, and operational efficiency of Industrial Internet of Things systems in smart city environments. However, existing federated learning based anomaly detection frameworks struggle to maintain high performance under dynamic operating conditions, non-independent data distributions, and heterogeneous edge device capabilities. Most current approaches lack effective mechanisms to adapt to concept drift and long-term temporal variations while preserving communication efficiency and data privacy. To address these limitations, this study proposes FLAME IoT. This adaptive federated learning framework integrates dynamic client utility modelling, momentum-based global model evolution, and hybrid temporal anomaly scoring for privacy-preserving real-time anomaly detection. The proposed framework assigns client participation weights based on data quality and resource availability, stabilises global learning through momentum-driven aggregation, and enhances detection sensitivity by combining reconstruction error with temporal dependency modelling. Theoretical convergence guarantees are established in heterogeneous, drift-prone environments. Extensive experiments are conducted across three benchmark Industrial IoT datasets: SWaT, TON IoT, and LoRaWAN. The results demonstrate that FLAME IoT achieves up to 18.7 per cent improvement in F1 Score and reduces communication overhead by approximately 32 per cent compared to conventional federated learning baselines. Additional evaluations confirm its robustness under non-IID data, synthetic concept drift, and large-scale client participation. Real-world testbed experiments further validate its practical effectiveness and energy efficiency. The main contribution of this work is a unified, adaptive federated anomaly detection framework that jointly addresses data heterogeneity, temporal dependency modelling, resource constraints, and evolving industrial conditions. By integrating adaptive learning, stable aggregation, and hybrid anomaly scoring into a single architecture, FLAME IoT provides a scalable, reliable solution for next-generation smart industrial systems.</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

FLAME-IoT: Federated Learning with Adaptive Model Evolution for Smart City Industrial IoT Anomaly Detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月19日

查看全文
  • Mohammed Naif Alatawi

摘要

Anomaly detection plays a vital role in ensuring the safety, reliability, and operational efficiency of Industrial Internet of Things systems in smart city environments. However, existing federated learning based anomaly detection frameworks struggle to maintain high performance under dynamic operating conditions, non-independent data distributions, and heterogeneous edge device capabilities. Most current approaches lack effective mechanisms to adapt to concept drift and long-term temporal variations while preserving communication efficiency and data privacy. To address these limitations, this study proposes FLAME IoT. This adaptive federated learning framework integrates dynamic client utility modelling, momentum-based global model evolution, and hybrid temporal anomaly scoring for privacy-preserving real-time anomaly detection. The proposed framework assigns client participation weights based on data quality and resource availability, stabilises global learning through momentum-driven aggregation, and enhances detection sensitivity by combining reconstruction error with temporal dependency modelling. Theoretical convergence guarantees are established in heterogeneous, drift-prone environments. Extensive experiments are conducted across three benchmark Industrial IoT datasets: SWaT, TON IoT, and LoRaWAN. The results demonstrate that FLAME IoT achieves up to 18.7 per cent improvement in F1 Score and reduces communication overhead by approximately 32 per cent compared to conventional federated learning baselines. Additional evaluations confirm its robustness under non-IID data, synthetic concept drift, and large-scale client participation. Real-world testbed experiments further validate its practical effectiveness and energy efficiency. The main contribution of this work is a unified, adaptive federated anomaly detection framework that jointly addresses data heterogeneity, temporal dependency modelling, resource constraints, and evolving industrial conditions. By integrating adaptive learning, stable aggregation, and hybrid anomaly scoring into a single architecture, FLAME IoT provides a scalable, reliable solution for next-generation smart industrial systems.