Scholarlink

<p>Network attack detection and prevention are critical aspects of ensuring cybersecurity in today’s interconnected world. Traditional centralized models for detecting malicious activity often face challenges like data privacy concerns and inefficiencies in handling distributed data sources. To address these challenges, this research leverages federated learning. Being a decentralized approach that offers advantageous benefits of collaborative model training across different devices while keeping sensitive data localized. Using the NF-UQ-NIDS-v2 benchmark dataset, we developed a robust machine-learning pipeline. This included thorough data preparation and applying advanced feature engineering techniques to reduce noise and enhance the dataset’s quality. We build several advanced machine learning approaches and hybrid models. Building on this, we proposed an innovative federated learning model designed to improve attack detection accuracy while preserving data privacy. With the help of the proposed federated learning architecture, the client model logistic regression (LR) achieved high-performance accuracy scores of 99% for multi-class network attack detection. Also, we quantify FL communication efficiency by measuring per-round latency, bandwidth usage, and the impact of compression techniques such as sparsification, 8-bit quantization, and top-<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(k\)</EquationSource> </InlineEquation> selection, reducing uplink payloads by up to 30<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>. This research uniquely harmonizes the NF-UQ-NIDS-v2 dataset by consolidating diverse attack classes into a unified schema, analyzes communication overhead in FL by quantifying latency, bandwidth usage, and synchronization costs under compression, and incorporates a heterogeneity-aware aggregation strategy that re-weights client contributions based on data skewness, thereby improving stability in non-IID settings. The proposed approach demonstrates significant potential for real-world applications, particularly in environments where data distribution and security are paramount.</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

Novel federated learning-based approach for network attack detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月10日

查看全文
  • Vinothkumar Kolluru,
  • Sudeep Mungara,
  • Ali Raza,
  • Ghazia Aslam,
  • Nagwan Abdel Samee,
  • Imran Ashraf

摘要

Network attack detection and prevention are critical aspects of ensuring cybersecurity in today’s interconnected world. Traditional centralized models for detecting malicious activity often face challenges like data privacy concerns and inefficiencies in handling distributed data sources. To address these challenges, this research leverages federated learning. Being a decentralized approach that offers advantageous benefits of collaborative model training across different devices while keeping sensitive data localized. Using the NF-UQ-NIDS-v2 benchmark dataset, we developed a robust machine-learning pipeline. This included thorough data preparation and applying advanced feature engineering techniques to reduce noise and enhance the dataset’s quality. We build several advanced machine learning approaches and hybrid models. Building on this, we proposed an innovative federated learning model designed to improve attack detection accuracy while preserving data privacy. With the help of the proposed federated learning architecture, the client model logistic regression (LR) achieved high-performance accuracy scores of 99% for multi-class network attack detection. Also, we quantify FL communication efficiency by measuring per-round latency, bandwidth usage, and the impact of compression techniques such as sparsification, 8-bit quantization, and top- \(k\) selection, reducing uplink payloads by up to 30 \(\times\) . This research uniquely harmonizes the NF-UQ-NIDS-v2 dataset by consolidating diverse attack classes into a unified schema, analyzes communication overhead in FL by quantifying latency, bandwidth usage, and synchronization costs under compression, and incorporates a heterogeneity-aware aggregation strategy that re-weights client contributions based on data skewness, thereby improving stability in non-IID settings. The proposed approach demonstrates significant potential for real-world applications, particularly in environments where data distribution and security are paramount.