A Novel Detection and Multi-Classification Approach for IoT-Malware Using Random Forest Voting of Fine-Tuning...

Area of Science:

• Cybersecurity
• Machine Learning
• Computer Vision

Background:

• Internet of Things (IoT) devices are increasingly vulnerable to malware due to inherent security weaknesses.
• Traditional IoT malware detection methods are insufficient for effective mitigation and prevention.
• Deep learning offers advanced capabilities for analyzing complex malware behavior.

Purpose of the Study:

• To propose a novel, vision-based deep learning approach for IoT malware detection and multi-classification.
• To leverage deep transfer learning, fine-tuning, and ensembling strategies to improve performance without training from scratch.
• To enhance the accuracy and efficiency of identifying IoT malware threats.

Main Methods:

• A fusion of three Convolutional Neural Networks (CNNs): ResNet18, MobileNetV2, and DenseNet161.
• Implementation of deep transfer learning and fine-tuning methodologies.
• Utilization of a random forest voting strategy for ensembling the CNN models.
• Validation using the MaleVis dataset comprising 14,226 images across 25 malware and 1 benign class.

Main Results:

• The proposed approach achieved high performance metrics: 98.74% precision, 98.67% recall, 98.79% specificity, 98.70% F1-score, 98.65% MCC, and 98.68% accuracy.
• Outperformed existing state-of-the-art solutions in IoT malware detection and classification.
• Demonstrated an average processing time of 672 ms per malware classification.

Conclusions:

• The developed vision-based deep learning model offers a highly effective solution for IoT malware detection and classification.
• The combination of deep transfer learning and CNN ensembling significantly boosts detection and classification performance.
• This approach provides a robust and efficient method for enhancing the security of Internet of Things devices against sophisticated malware threats.