Scholarlink

<p>Federated learning (FL) has gained prominence as a decentralized approach for training machine learning models on distributed data while safeguarding user privacy. However, challenges arise from the non-independent and identically distributed (non-IID) nature of network traffic data and the necessity for secure model aggregation in heterogeneous environments. To address these issues, we propose AFL-SecNet, an advanced FL framework that dynamically adapts local training processes to accommodate diverse network traffic types across clients, thereby improving model performance and generalization in dynamic settings. AFL-SecNet incorporates a distributed two-trapdoor public-key cryptosystem for secure aggregation, ensuring that model updates remain encrypted and require dual-party decryption to prevent unauthorized access. Furthermore, it enables client-specific personalization by fine-tuning the global model on local datasets, improving accuracy while preserving the benefits of collaborative learning. We evaluate AFL-SecNet on the UNSW-NB15 and CICDDoS2019 datasets, encompassing normal traffic, DDoS attacks, and anomalies. On UNSW-NB15, the model achieves 97.58% accuracy, 98.29% recall, 96.72% precision, 97.81% F1-score, 0.0315 MSE, 0.0391% FAR, and 1.15 sec inference time, showing strong intrusion detection. On CICDDoS2019, it attains 96.60% accuracy, 95.77% recall, 94.22% precision, 96.81% F1-score, 0.0367 MSE, 0.0421% FAR, and 0.63 sec inference time, demonstrating robustness in DDoS classification. These outcomes underscore AFL-SecNet’s capability to manage non-IID data, bolster privacy via robust cryptography, and elevate detection performance through personalization, advancing scalable and secure FL solutions for network traffic analysis.</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

AFL-SecNet: An adaptive federated learning framework for secure and privacy-preserving network traffic analysis

  • 期刊论文

  • 发表日期：

    2025年12月11日

查看全文
  • Abida Naz,
  • Irshad Ullah,
  • Muhammad Uzair,
  • Muhammad Fayyaz Khokhar,
  • Ayesha Sabir,
  • Riaz Ullah Khan

摘要

Federated learning (FL) has gained prominence as a decentralized approach for training machine learning models on distributed data while safeguarding user privacy. However, challenges arise from the non-independent and identically distributed (non-IID) nature of network traffic data and the necessity for secure model aggregation in heterogeneous environments. To address these issues, we propose AFL-SecNet, an advanced FL framework that dynamically adapts local training processes to accommodate diverse network traffic types across clients, thereby improving model performance and generalization in dynamic settings. AFL-SecNet incorporates a distributed two-trapdoor public-key cryptosystem for secure aggregation, ensuring that model updates remain encrypted and require dual-party decryption to prevent unauthorized access. Furthermore, it enables client-specific personalization by fine-tuning the global model on local datasets, improving accuracy while preserving the benefits of collaborative learning. We evaluate AFL-SecNet on the UNSW-NB15 and CICDDoS2019 datasets, encompassing normal traffic, DDoS attacks, and anomalies. On UNSW-NB15, the model achieves 97.58% accuracy, 98.29% recall, 96.72% precision, 97.81% F1-score, 0.0315 MSE, 0.0391% FAR, and 1.15 sec inference time, showing strong intrusion detection. On CICDDoS2019, it attains 96.60% accuracy, 95.77% recall, 94.22% precision, 96.81% F1-score, 0.0367 MSE, 0.0421% FAR, and 0.63 sec inference time, demonstrating robustness in DDoS classification. These outcomes underscore AFL-SecNet’s capability to manage non-IID data, bolster privacy via robust cryptography, and elevate detection performance through personalization, advancing scalable and secure FL solutions for network traffic analysis.