Scholarlink

<p>The convergence of smart grids, electric vehicles (EVs), and Vehicular Ad Hoc Network (VANET) infrastructures has created a rapidly evolving cyber-physical ecosystem that necessitates real-time, privacy-preserving, and intelligent threat detection at the edge. It is crucial to develop an unified threat intelligence across these diverse domains to ensure operational resilience and data integrity. Existing threat detection approaches frequently depend on centralized training pipelines and isolated domain-specific models, resulting in substantial communication overhead, poor scalability, and reduced robustness under domain shift.&#xa0;This ultimately leads to very low detection accuracy when deploying these models in heterogeneous environments with different data distributions. To address these challenges, we propose a Federated Lightweight Cross-Domain Adversarial Domain Adaptation (FL-CDA) framework. FL-CDA enables edge nodes to collaboratively train compact models via federated learning while incorporating adversarial domain adaptation to align heterogeneous data distributions without the need to share raw data. The proposed framework is designed to support the detection of false data injection in Smart Grids, EV charging fraud, and Sybil or spoofing attacks in VANETs while preserving data privacy and reducing bandwidth usage. Experiments in a hybrid evaluation setting comprising a real-world smart-grid dataset and synthetic EV/VANET scenarios indicate that FL-CDA significantly improves cross-domain detection accuracy, reduces communication overhead, and enhances model robustness under dynamic adversarial conditions. The experiments show up to 95% cross-domain accuracy, 1.9 s detection latency, 94% F1-score, and 25–110 MB communication overhead in the simulated setting, while sustaining 86% accuracy with a 50% domain shift.</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

Cross-domain edge AI framework for unified threat intelligence in smart grid-EV- VANET ecosystems using lightweight federated learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月14日

查看全文
  • Subbiah Swaminathan,
  • T. Avudaiappan,
  • B. Baazeer Ahamed,
  • Emerson Raja Joseph

摘要

The convergence of smart grids, electric vehicles (EVs), and Vehicular Ad Hoc Network (VANET) infrastructures has created a rapidly evolving cyber-physical ecosystem that necessitates real-time, privacy-preserving, and intelligent threat detection at the edge. It is crucial to develop an unified threat intelligence across these diverse domains to ensure operational resilience and data integrity. Existing threat detection approaches frequently depend on centralized training pipelines and isolated domain-specific models, resulting in substantial communication overhead, poor scalability, and reduced robustness under domain shift. This ultimately leads to very low detection accuracy when deploying these models in heterogeneous environments with different data distributions. To address these challenges, we propose a Federated Lightweight Cross-Domain Adversarial Domain Adaptation (FL-CDA) framework. FL-CDA enables edge nodes to collaboratively train compact models via federated learning while incorporating adversarial domain adaptation to align heterogeneous data distributions without the need to share raw data. The proposed framework is designed to support the detection of false data injection in Smart Grids, EV charging fraud, and Sybil or spoofing attacks in VANETs while preserving data privacy and reducing bandwidth usage. Experiments in a hybrid evaluation setting comprising a real-world smart-grid dataset and synthetic EV/VANET scenarios indicate that FL-CDA significantly improves cross-domain detection accuracy, reduces communication overhead, and enhances model robustness under dynamic adversarial conditions. The experiments show up to 95% cross-domain accuracy, 1.9 s detection latency, 94% F1-score, and 25–110 MB communication overhead in the simulated setting, while sustaining 86% accuracy with a 50% domain shift.