BlockDroid: detection of Android malware from images using lightweight convolutional neural network models with...

Area of Science:

• Computer Science
• Cybersecurity
• Machine Learning

Background:

• Traditional Android malware detection methods struggle with increasing malware complexity and evasion techniques.
• Signature-based detection is insufficient against polymorphic and obfuscated malware.
• Machine learning, particularly deep learning, offers advanced capabilities for automated feature extraction and complex pattern recognition in malware detection.

Purpose of the Study:

• To introduce BlockDroid, an innovative approach for accurate and efficient Android malware detection.
• To leverage convolutional neural network (CNN) models, ensemble learning, and blockchain technology for enhanced detection capabilities.
• To address the limitations of traditional methods and improve resource utilization in mobile security.

Main Methods:

• Android DEX files were converted into image data for analysis by CNN models.
• Lightweight CNN models (EfficientNetB0, MobileNetV2, custom) were employed.
• Ensemble learning and blockchain technology were integrated for improved accuracy and efficient record-keeping.

Main Results:

• The BlockDroid system achieved a high accuracy rate of 97.38% on the CICMalDroid 2020 dataset.
• BlockDroid outperformed individual CNN models in malware detection accuracy.
• Blockchain integration enabled efficient recording of predictions, avoiding redundant analysis.

Conclusions:

• BlockDroid presents a highly accurate and efficient solution for Android malware detection.
• The combination of deep learning and blockchain technology offers a novel and effective approach to mobile security.
• This method significantly improves resource utilization by preventing repeated analysis of applications.