Scholarlink

<p>The integration of Internet of Things (IoT) systems with green cloud infrastructures enhances scalability and energy efficiency but simultaneously increases exposure to adaptive phishing threats. Conventional signature-based and static supervised learning methods lack adaptability and fail to address the strict energy constraints of decentralized IoT deployments. To overcome these limitations, this paper proposes HEAP-Guard, a Hybrid Energy-Aware Phishing Detection framework built upon a Reinforcement Learning (RL) architecture augmented with Risk- and Energy-Aware optimization objectives. To enhance training stability, the proposed REA-TDX mechanism extends the RL backbone through twin critics, delayed policy updates, and conservative Bellman backups, ensuring robust and stable convergence. A hybrid execution–training architecture further enables lightweight deterministic inference at the network edge while preserving stability-enhanced policy learning within the cloud. Moreover, a FedAvg-driven collaborative training strategy allows distributed IoT nodes to jointly optimize shared policies without exchanging raw traffic data, thereby improving generalization capability and preserving privacy. Unlike static classifiers, HEAP-Guard produces continuous phishing confidence scores and jointly optimizes detection performance, operational risk, and energy consumption. Experimental results on the Bot-IoT dataset demonstrate that the proposed model achieves 99.32% ± 0.18 accuracy, with 97.9% precision, 97.3% recall, and 97.6% F1-score, outperforming classical Machine Learning, Deep Learning and baseline RL approaches. In addition, the framework reduces per-inference energy consumption by up to 38.7% compared to CNN baselines, operating at approximately 1.65&#xa0;W in balanced edge mode, thereby confirming its suitability for energy-efficient IoT-enabled green cloud environments.</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

HEAP-Guard: A hybrid energy-aware phishing detection framework based on risk and stability-aware reinforcement learning for IoT green clouds

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Firas Saidi,
  • Zied Ben Hazem,
  • Nivine Guler

摘要

The integration of Internet of Things (IoT) systems with green cloud infrastructures enhances scalability and energy efficiency but simultaneously increases exposure to adaptive phishing threats. Conventional signature-based and static supervised learning methods lack adaptability and fail to address the strict energy constraints of decentralized IoT deployments. To overcome these limitations, this paper proposes HEAP-Guard, a Hybrid Energy-Aware Phishing Detection framework built upon a Reinforcement Learning (RL) architecture augmented with Risk- and Energy-Aware optimization objectives. To enhance training stability, the proposed REA-TDX mechanism extends the RL backbone through twin critics, delayed policy updates, and conservative Bellman backups, ensuring robust and stable convergence. A hybrid execution–training architecture further enables lightweight deterministic inference at the network edge while preserving stability-enhanced policy learning within the cloud. Moreover, a FedAvg-driven collaborative training strategy allows distributed IoT nodes to jointly optimize shared policies without exchanging raw traffic data, thereby improving generalization capability and preserving privacy. Unlike static classifiers, HEAP-Guard produces continuous phishing confidence scores and jointly optimizes detection performance, operational risk, and energy consumption. Experimental results on the Bot-IoT dataset demonstrate that the proposed model achieves 99.32% ± 0.18 accuracy, with 97.9% precision, 97.3% recall, and 97.6% F1-score, outperforming classical Machine Learning, Deep Learning and baseline RL approaches. In addition, the framework reduces per-inference energy consumption by up to 38.7% compared to CNN baselines, operating at approximately 1.65 W in balanced edge mode, thereby confirming its suitability for energy-efficient IoT-enabled green cloud environments.