Scholarlink

The exponential rise in Android ransomware attacks poses severe security threats, leading to financial losses and privacy breaches for both individuals and businesses. Traditional signature-based detection techniques struggle against evolving ransomware variants, necessitating more advanced and adaptive detection mechanisms. This work proposes an approach that first employs Random Forest classification for structured feature-based analysis. To further enhance detection, JSON reports are transformed into images, and deep learning models, including Convolutional Neural Network (CNN) and Vision Transformer (ViT), are applied for ransomware classification. We extracted behavioral reports from the CuckooDroid sandbox, collecting approximately 2280 ransomware reports and 2000 benign application reports. A Random Forest classifier trained on structured features achieves an accuracy of 99.41%. To further enhance Android ransomware detection, we introduce a novel transformation of JSON reports into RGB images, enabling deep-learning models to classify ransomware patterns. Our CNN trained on RGB images attains 99.76% accuracy, while a CNN trained on grayscale images achieves 99.53%. Further, the ViT model trained on RGB images optimized using the grid search surpasses all models, achieving a peak accuracy of 99.78%.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Behavior to Pixels: A Vision Transformer Approach for Android Ransomware Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Satyam Kesharwani,
  • Kamaldeep,
  • Manisha Malik

摘要

The exponential rise in Android ransomware attacks poses severe security threats, leading to financial losses and privacy breaches for both individuals and businesses. Traditional signature-based detection techniques struggle against evolving ransomware variants, necessitating more advanced and adaptive detection mechanisms. This work proposes an approach that first employs Random Forest classification for structured feature-based analysis. To further enhance detection, JSON reports are transformed into images, and deep learning models, including Convolutional Neural Network (CNN) and Vision Transformer (ViT), are applied for ransomware classification. We extracted behavioral reports from the CuckooDroid sandbox, collecting approximately 2280 ransomware reports and 2000 benign application reports. A Random Forest classifier trained on structured features achieves an accuracy of 99.41%. To further enhance Android ransomware detection, we introduce a novel transformation of JSON reports into RGB images, enabling deep-learning models to classify ransomware patterns. Our CNN trained on RGB images attains 99.76% accuracy, while a CNN trained on grayscale images achieves 99.53%. Further, the ViT model trained on RGB images optimized using the grid search surpasses all models, achieving a peak accuracy of 99.78%.