Machine learning models and dimensionality reduction for improving the Android malware detection DOI Creative Commons

P.R. Moran,

Antonio Robles-Gómez, Andrés Duque

et al.

PeerJ Computer Science, Journal Year: 2024, Volume and Issue: 10, P. e2616 - e2616

Published: Dec. 23, 2024

Today, a great number of attack opportunities for cybercriminals arise in Android, since it is one the most used operating systems many mobile applications. Hence, very important to anticipate these situations. To minimize this problem, analysis malware search applications based on machine learning algorithms. Our work uses as starting point features proposed by DREBIN project, which today constitutes key reference literature, being largest public Android dataset with labeled families. The authors only employ support vector determine whether sample or not. This first proposes new efficient dimensionality reduction features, well application several supervised algorithms prediction purposes. Predictive models Random Forest are found achieve promising results. They can detect an average 91.72% samples, low false positive rate 0.13%, and using 5,000 features. just over 9% total DREBIN. It achieves accuracy 99.52%, precision 96.91%, macro F1-score 96.99%.

Language: Английский

Detecting Obfuscated Malware Infections on Windows Using Ensemble Learning Techniques DOI Creative Commons
Yadigar İmamverdiyev, Elshan Baghirov,

John Chukwu Ikechukwu

et al.

Informatics and Automation, Journal Year: 2025, Volume and Issue: 24(1), P. 99 - 124

Published: Jan. 20, 2025

In the internet and smart devices era, malware detection has become crucial for system security. Obfuscated poses significant risks to various platforms, including computers, mobile devices, IoT by evading advanced security solutions. Traditional heuristic-based signature-based methods often fail against these threats. Therefore, a cost-effective was proposed using memory dump analysis ensemble learning techniques. Utilizing CIC-MalMem-2022 dataset, effectiveness of decision trees, gradient-boosted logistic Regression, random forest, LightGBM in identifying obfuscated evaluated. The study demonstrated superiority techniques enhancing accuracy robustness. Additionally, SHAP (SHapley Additive exPlanations) LIME (Local Interpretable Model-agnostic Explanations) were employed elucidate model predictions, improving transparency trustworthiness. revealed vital features significantly impacting detection, such as process services, active file handles, registry keys, callback functions. These insights are refining strategies performance. findings contribute cybersecurity efforts comprehensively assessing machine algorithms through analysis. This paper offers valuable future research advancements paving way more robust effective solutions face evolving sophisticated

Language: Английский

Citations

0
Machine learning models and dimensionality reduction for improving the Android malware detection DOI Creative Commons

P.R. Moran,

Antonio Robles-Gómez, Andrés Duque

et al.

PeerJ Computer Science, Journal Year: 2024, Volume and Issue: 10, P. e2616 - e2616

Published: Dec. 23, 2024

Today, a great number of attack opportunities for cybercriminals arise in Android, since it is one the most used operating systems many mobile applications. Hence, very important to anticipate these situations. To minimize this problem, analysis malware search applications based on machine learning algorithms. Our work uses as starting point features proposed by DREBIN project, which today constitutes key reference literature, being largest public Android dataset with labeled families. The authors only employ support vector determine whether sample or not. This first proposes new efficient dimensionality reduction features, well application several supervised algorithms prediction purposes. Predictive models Random Forest are found achieve promising results. They can detect an average 91.72% samples, low false positive rate 0.13%, and using 5,000 features. just over 9% total DREBIN. It achieves accuracy 99.52%, precision 96.91%, macro F1-score 96.99%.

Language: Английский

Citations

1