Scholarlink

<p>Buffer overflow is a severe software vulnerability threatening system security and reliability. This work presents a novel detection model leveraging token-level semantic embeddings from GraphCodeBERT combined with structural Program Dependence Graph (PDG) features. A Gated Attention Fusion mechanism effectively integrates semantic and structural information, enabling precise identification of buffer misuse patterns. The hybrid deep architecture, incorporating Convolutional Neural Networks (CNN) and Artificial Neural Networks (ANN), captures both local and global code features. To optimize performance and reduce manual tuning, Hybrid Manta Ray Bee Optimization (HMRBO) is employed for hyperparameter tuning. Extensive benchmarking demonstrates that the proposed method achieves superior detection accuracy and F1-score compared to state-of-the-art approaches, with interpretable fault localization via attention visualization. This study advances automated buffer overflow detection by combining deep code embeddings, structural graph analysis, and bio-inspired optimization techniques.</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

Buffer overflow fault detection using HMRBO, CNN-ANN with GraphcodeBERT, PDG features, and gated attention fusion

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • S. Thenmozhi,
  • PM. Shanthi

摘要

Buffer overflow is a severe software vulnerability threatening system security and reliability. This work presents a novel detection model leveraging token-level semantic embeddings from GraphCodeBERT combined with structural Program Dependence Graph (PDG) features. A Gated Attention Fusion mechanism effectively integrates semantic and structural information, enabling precise identification of buffer misuse patterns. The hybrid deep architecture, incorporating Convolutional Neural Networks (CNN) and Artificial Neural Networks (ANN), captures both local and global code features. To optimize performance and reduce manual tuning, Hybrid Manta Ray Bee Optimization (HMRBO) is employed for hyperparameter tuning. Extensive benchmarking demonstrates that the proposed method achieves superior detection accuracy and F1-score compared to state-of-the-art approaches, with interpretable fault localization via attention visualization. This study advances automated buffer overflow detection by combining deep code embeddings, structural graph analysis, and bio-inspired optimization techniques.