Scholarlink

<p>For mitigating the security breaches in FL based IDSs for Industrial IoT (IIoT) and to optimize the negative impact on model performance of Differential Privacy (DP) to prevent data leakages. In this paper, authors propose a novel approach of DP which is utilized at the time of training in FL for IIoT. Contextual Anonymization Aware Differential Privacy (CtxADP) is a mechanism applied by extending the concept of t-closeness in a context-sensitive manner and employed of DP once anonymization has been introduced using several parameters that optimize the privacy and utility tradeoff in the data. In addition to this, authors utilized the principle of Active Learning (ALrn) and Meta Learning (MLrn) in a Federated context, targeting to enhance the effectiveness and adaptability of the learning procedure. Moreover, to removes single points of failure, ensuring trust among IIoT devices without requiring a central authority. Authors employ Blockchain to enhance security in Federated Active Meta Learning (FedAMLrn) providing decentralized validation, immutability, and tamper-proof model updates. Unlike traditional FL, which relies on a central aggregator, Blockchain enables distributed consensus using Byzantine Fault Tolerance (ByzFT), ensuring that only legitimate model updates are accepted. Each model update is hashed and recorded in an immutable ledger, preventing adversarial modifications. Furthermore, authors generate evolving Actionable Intelligence Graph (AIG) that visualize detected threats and deliver insights for security administrators. Authors conducted experiments with <b>CIC-ToN-IoT, Bot-IoT</b> and <b>UNSW-NB15</b> datasets across the proposed approach <b>FedAMLrn</b> that shows the significant trade-off between privacy and utility. Experimental results indicate that FedAMLrn achieves an accuracy of 96.7 percent, surpassing conventional FL 91.5 percent, MTL 90.5 percent, and DistL 89 percent, while maintaining the lowest communication overhead 5.5&#xa0;MB and highest privacy preservation score 0.89 with BMC.</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 active meta-learning with blockchain for zero-day attack detection in industrial IoT

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Kapil Kumar,
  • Manju Khari

摘要

For mitigating the security breaches in FL based IDSs for Industrial IoT (IIoT) and to optimize the negative impact on model performance of Differential Privacy (DP) to prevent data leakages. In this paper, authors propose a novel approach of DP which is utilized at the time of training in FL for IIoT. Contextual Anonymization Aware Differential Privacy (CtxADP) is a mechanism applied by extending the concept of t-closeness in a context-sensitive manner and employed of DP once anonymization has been introduced using several parameters that optimize the privacy and utility tradeoff in the data. In addition to this, authors utilized the principle of Active Learning (ALrn) and Meta Learning (MLrn) in a Federated context, targeting to enhance the effectiveness and adaptability of the learning procedure. Moreover, to removes single points of failure, ensuring trust among IIoT devices without requiring a central authority. Authors employ Blockchain to enhance security in Federated Active Meta Learning (FedAMLrn) providing decentralized validation, immutability, and tamper-proof model updates. Unlike traditional FL, which relies on a central aggregator, Blockchain enables distributed consensus using Byzantine Fault Tolerance (ByzFT), ensuring that only legitimate model updates are accepted. Each model update is hashed and recorded in an immutable ledger, preventing adversarial modifications. Furthermore, authors generate evolving Actionable Intelligence Graph (AIG) that visualize detected threats and deliver insights for security administrators. Authors conducted experiments with CIC-ToN-IoT, Bot-IoT and UNSW-NB15 datasets across the proposed approach FedAMLrn that shows the significant trade-off between privacy and utility. Experimental results indicate that FedAMLrn achieves an accuracy of 96.7 percent, surpassing conventional FL 91.5 percent, MTL 90.5 percent, and DistL 89 percent, while maintaining the lowest communication overhead 5.5 MB and highest privacy preservation score 0.89 with BMC.