Scholarlink

<p>With the rapid evolution of cyber threats, traditional security frameworks struggle to keep pace with increasingly sophisticated attack patterns, which are often designed to evade conventional detection mechanisms. Cyber attackers continuously develop new techniques, including polymorphic malware, advanced persistent threats (APTs), and AI-driven adversarial attacks, making it difficult for static rule-based or signature-based intrusion detection systems (IDS) to identify novel threats effectively. Additionally, modern network environments generate vast amounts of high-dimensional traffic data, comprising diverse communication protocols, encrypted packets, and complex attack signatures, which can overwhelm traditional security models and lead to inefficiencies in real-time analysis. This research presents an advanced cyber-attack detection framework leveraging Transformers-VGRH (Variable Graph Representation with Hierarchies), a deep learning architecture designed to model complex attack behaviors using attention-based feature extraction. The model is further optimized using a hybrid Gray Wolf Optimization (GWO) and particle swarm optimization (PSO) approach, ensuring precise hyperparameter tuning for enhanced detection accuracy and reduced computational complexity. The proposed method is evaluated on two benchmark cybersecurity datasets: CICIDS-2017 and TON-IoT, which contain real-world attack scenarios, including network intrusions and IoT-based cyber threats. To handle high-dimensional data, principal component analysis (PCA) is applied for feature reduction, while Synthetic Minority Over-Sampling Technique (SMOTE) addresses dataset imbalance by generating synthetic attack samples. Additionally, Big Bang Big Crunch (BBBC) is utilized for optimal feature selection, improving model performance. The GWO-PSO-Transformers-VGRH framework achieves exceptional detection performance, with an accuracy of 98.7%, precision of 97.5%, recall of 97.8%, and an F1-score of 97.6%, significantly outperforming conventional approaches. The integration of graph-based deep learning, attention mechanisms, and hybrid optimization offers a scalable and robust solution for real-time cyber-attack detection in both traditional network infrastructures and IoT environments, making it highly applicable for enterprise security and large-scale cybersecurity monitoring.</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

Deep neural network-based cyber-attack avoidance system with hybrid optimization

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Prashant Kumar Shukla,
  • Ratish Agarwal

摘要

With the rapid evolution of cyber threats, traditional security frameworks struggle to keep pace with increasingly sophisticated attack patterns, which are often designed to evade conventional detection mechanisms. Cyber attackers continuously develop new techniques, including polymorphic malware, advanced persistent threats (APTs), and AI-driven adversarial attacks, making it difficult for static rule-based or signature-based intrusion detection systems (IDS) to identify novel threats effectively. Additionally, modern network environments generate vast amounts of high-dimensional traffic data, comprising diverse communication protocols, encrypted packets, and complex attack signatures, which can overwhelm traditional security models and lead to inefficiencies in real-time analysis. This research presents an advanced cyber-attack detection framework leveraging Transformers-VGRH (Variable Graph Representation with Hierarchies), a deep learning architecture designed to model complex attack behaviors using attention-based feature extraction. The model is further optimized using a hybrid Gray Wolf Optimization (GWO) and particle swarm optimization (PSO) approach, ensuring precise hyperparameter tuning for enhanced detection accuracy and reduced computational complexity. The proposed method is evaluated on two benchmark cybersecurity datasets: CICIDS-2017 and TON-IoT, which contain real-world attack scenarios, including network intrusions and IoT-based cyber threats. To handle high-dimensional data, principal component analysis (PCA) is applied for feature reduction, while Synthetic Minority Over-Sampling Technique (SMOTE) addresses dataset imbalance by generating synthetic attack samples. Additionally, Big Bang Big Crunch (BBBC) is utilized for optimal feature selection, improving model performance. The GWO-PSO-Transformers-VGRH framework achieves exceptional detection performance, with an accuracy of 98.7%, precision of 97.5%, recall of 97.8%, and an F1-score of 97.6%, significantly outperforming conventional approaches. The integration of graph-based deep learning, attention mechanisms, and hybrid optimization offers a scalable and robust solution for real-time cyber-attack detection in both traditional network infrastructures and IoT environments, making it highly applicable for enterprise security and large-scale cybersecurity monitoring.