Scholarlink

<p>The sheer growth of smart agricultural systems in the Agriculture 4.0 paradigm has presented a growing attack surface that traditional intrusion detection methods are ill-equipped to detect. IoT gateway devices located in remote farm environments are constrained by resources, and are vulnerable to advanced adversarial threats such as coordinated distributed denial-of-service (DDoS) attacks, or insidious advanced persistent threats (APTs). This paper introduces FedTrans-AgriIDS, a new architecture that combines federated deep reinforcement learning (FedDRL) with a Transformer-based anomaly detector (TransAD) to provide privacy-aware, adaptable, and low-resource intrusion detection in agricultural IoT networks. The FedDRL aspect uses a multi-agent proximal policy optimization (MAPPO) approach that spreads the learning load across a heterogeneous set of edge gateways without broadcasting raw sensor measurements, thus ensuring the sovereignty of agronomic data. TransAD uses multi-head self-attention and positional encoding to learn long-term temporal dependencies between network traffic sequences to detect both zero-day exploits and low-rate DDoS variants, which subvert signature-based tools. Three testbeds were experimented, including a physical Raspberry Pi 4 cluster, a virtualized edge-environment, which mimicked the LPWAN and NB-IoT connectivity, and a benchmark comparison to the CIC-IoT-2023 and UNSW-NB15 datasets. On sustained attack campaigns, FedTrans-AgriIDS had an overall detection accuracy of 98.73, F1-score of 0.9861 and a false-positive rate of 0.41% with limited CPU usage (35 percent) and memory usage (22 percent). The proposed framework yielded lower false negatives on APT-class threats by 64.2 compared to standalone Zeek IDS, Snort 3, and a centralized LSTM-based detector and found shorter mean detection latencies by 38.7&#xa0;ms. these results confirm that FedTrans-AgriIDS is a viable and scalable security layer that can be deployed in a variety of smart farming infrastructures.</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

Federated deep reinforcement learning with transformer-based anomaly detection for cybersecurity in next-generation agricultural IoT networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月25日

查看全文
  • P. Sundaravadivel,
  • R. Augustian Isaac,
  • P. Vijayan

摘要

The sheer growth of smart agricultural systems in the Agriculture 4.0 paradigm has presented a growing attack surface that traditional intrusion detection methods are ill-equipped to detect. IoT gateway devices located in remote farm environments are constrained by resources, and are vulnerable to advanced adversarial threats such as coordinated distributed denial-of-service (DDoS) attacks, or insidious advanced persistent threats (APTs). This paper introduces FedTrans-AgriIDS, a new architecture that combines federated deep reinforcement learning (FedDRL) with a Transformer-based anomaly detector (TransAD) to provide privacy-aware, adaptable, and low-resource intrusion detection in agricultural IoT networks. The FedDRL aspect uses a multi-agent proximal policy optimization (MAPPO) approach that spreads the learning load across a heterogeneous set of edge gateways without broadcasting raw sensor measurements, thus ensuring the sovereignty of agronomic data. TransAD uses multi-head self-attention and positional encoding to learn long-term temporal dependencies between network traffic sequences to detect both zero-day exploits and low-rate DDoS variants, which subvert signature-based tools. Three testbeds were experimented, including a physical Raspberry Pi 4 cluster, a virtualized edge-environment, which mimicked the LPWAN and NB-IoT connectivity, and a benchmark comparison to the CIC-IoT-2023 and UNSW-NB15 datasets. On sustained attack campaigns, FedTrans-AgriIDS had an overall detection accuracy of 98.73, F1-score of 0.9861 and a false-positive rate of 0.41% with limited CPU usage (35 percent) and memory usage (22 percent). The proposed framework yielded lower false negatives on APT-class threats by 64.2 compared to standalone Zeek IDS, Snort 3, and a centralized LSTM-based detector and found shorter mean detection latencies by 38.7 ms. these results confirm that FedTrans-AgriIDS is a viable and scalable security layer that can be deployed in a variety of smart farming infrastructures.