RansomDroid: Android Ransomware Detection

Researched and developed a novel deep learning approach for Android ransomware detection, achieving 99.78% accuracy using a Vision Transformer (ViT) and dynamic behavioral analysis.

Read the full write-up

6
Total Technologies
4
Key Features
Read Paper
Github

Technologies Used

Python
Vision Transformer
CNN
Machine Learning
CuckooDroid
Feature Engineering
RansomDroid: Android Ransomware Detection

Key Features

  • Vision Transformer: Fine-tuned a pretrained ViT for Android ransomware detection, using grid search over the learning rate and batch size, reaching 99.78% accuracy.
  • Diverse Data Creation: Executed dynamic analysis on a comprehensive dataset of 4,280 APKs (2,280 RansomProber ransomware and 2,000 Androzoo benign apps) through the CuckooDroid sandbox.
  • Customized Data Transformation: Introduced a novel transformation strategy converting CuckooDroid JSON reports into structured CSVs and visual formats (RGB & grayscale images) to capture subtle spatial behavioral patterns.
  • Comparative Evaluation: Compared the ViT against a CNN (99.76% on RGB and 99.53% on grayscale images) and a Random Forest (99.41% on CSV data).