Scholarlink

The growing volume and complexity of malware families pose significant cybersecurity threats, necessitating advanced techniques for accurate classification and detection. As cyber threats become increasingly sophisticated, protecting sensitive data, preventing financial losses, and ensuring digital security require improved detection methods. Traditional machine-learning approaches need to be revised for the evolving malware landscape, which involves obfuscation techniques and dynamic behaviors. Therefore, enhancing malware detection systems’ accuracy, efficiency, and robustness is crucial. Current malware detection techniques often rely on static analysis, which struggles with obfuscated and dynamic malware. This paper addresses gaps in the existing literature by focusing on advanced ensemble models, systematic hyperparameter optimization, and improved feature handling to enhance detection performance. Ensemble learning techniques, such as Gradient Boosting Machine (GBM), XGBoost, CatBoost, Random Forest, and LightGBM, offer improved accuracy, reduced overfitting, and increased robustness. These models also address issues with imbalanced data and enhance interpretability. The study utilizes a 0.5 terabyte dataset from the Microsoft Malware Classification Challenge 2015, containing nine malware families with unique identifiers. Models were trained and optimized using GridSearchCV to identify the best parameter values. The paper discusses challenges such as the need for multiple datasets, feature encoding issues, and handling dynamic features. Solutions proposed include applying feature selection algorithms, balancing sample distributions, and optimizing through parallelization. Preliminary investigations into advanced encoding techniques and ensemble learning models indicate that LightGBM performs optimally with an accuracy of 96.71%. The study also highlights the potential of transfer learning to enhance malware detection by leveraging prior knowledge from trained models. After optimization with machine learning algorithms, deep learning models demonstrated superior performance compared to traditional methods. This research revealed transfer learning and deep learning to be very promising for improving malware detection capabilities and achieving levels of efficiency and precision previously unattained.

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

Transfer Learning and Ensemble Models for Robust Malware Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Malay Doshi,
  • Gauri Deoghare,
  • Ranjeet Vasant Bidwe,
  • Sashikala Mishra

摘要

The growing volume and complexity of malware families pose significant cybersecurity threats, necessitating advanced techniques for accurate classification and detection. As cyber threats become increasingly sophisticated, protecting sensitive data, preventing financial losses, and ensuring digital security require improved detection methods. Traditional machine-learning approaches need to be revised for the evolving malware landscape, which involves obfuscation techniques and dynamic behaviors. Therefore, enhancing malware detection systems’ accuracy, efficiency, and robustness is crucial. Current malware detection techniques often rely on static analysis, which struggles with obfuscated and dynamic malware. This paper addresses gaps in the existing literature by focusing on advanced ensemble models, systematic hyperparameter optimization, and improved feature handling to enhance detection performance. Ensemble learning techniques, such as Gradient Boosting Machine (GBM), XGBoost, CatBoost, Random Forest, and LightGBM, offer improved accuracy, reduced overfitting, and increased robustness. These models also address issues with imbalanced data and enhance interpretability. The study utilizes a 0.5 terabyte dataset from the Microsoft Malware Classification Challenge 2015, containing nine malware families with unique identifiers. Models were trained and optimized using GridSearchCV to identify the best parameter values. The paper discusses challenges such as the need for multiple datasets, feature encoding issues, and handling dynamic features. Solutions proposed include applying feature selection algorithms, balancing sample distributions, and optimizing through parallelization. Preliminary investigations into advanced encoding techniques and ensemble learning models indicate that LightGBM performs optimally with an accuracy of 96.71%. The study also highlights the potential of transfer learning to enhance malware detection by leveraging prior knowledge from trained models. After optimization with machine learning algorithms, deep learning models demonstrated superior performance compared to traditional methods. This research revealed transfer learning and deep learning to be very promising for improving malware detection capabilities and achieving levels of efficiency and precision previously unattained.