The proliferation of malicious software poses a persistent threat to digital security, particularly in environments with constrained computational resources. Traditional malware detection systems often rely on resource-intensive operations, making them unsuitable for low-power devices. This research proposes an efficient machine learning-based approach for malware classification that emphasizes feature influence techniques to optimize performance for such devices. The study focuses on distinguishing between malware and legitimate software using a binary classification model. A comparative evaluation of existing algorithms Random Forest and Extra Trees classifiers is conducted against the proposed Decision Tree and XGBoost models. The goal is to assess model performance in terms of accuracy, precision, recall, and computational efficiency while reducing model complexity without compromising detection capabilities. Feature importance analysis is employed to select the most impactful attributes, thereby streamlining the learning process and minimizing resource usage. The experimental results demonstrate that the proposed models, particularly XGBoost, achieve competitive accuracy with significantly reduced computational overhead compared to ensemble-based methods. This confirms the effectiveness of employing lightweight classifiers alongside feature influence techniques in improving malware detection performance on resource-limited platforms. The findings suggest that intelligent model selection and targeted feature reduction are key strategies in developing practical and deployable malware detection systems suitable for devices with limited processing power.

Keywords: Malware detection, Machine learning, XGBoost, Decision tree, Feature selection, Binary classification, Lightweight models, Resource-constrained devices, Feature importance, Cybersecurity.