Scholarlink

QR code-based phishing (“quishing”) is a growing cybersecurity threat. The baseline XGBoost model by Trad and Chehab (2025) [1] achieves a ROC-AUC of 0.9133 on 69x69 images. We propose a feature learning pipeline integrating image-based, metadata, and optional URL-based features, evaluated with Vanilla CNN, Residual CNN, and XGBoost models. Targeting a ROC-AUC above the baseline, our Residual CNN achieves 0.9313, followed by XGBoost (0.9158) and Vanilla CNN (0.8900), with superior robustness to perturbations like blur and compression. Contributions include an optimized feature pipeline for low-resolution QR codes and a comparison of deep learning and boosting models for real-world quishing detection.

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

Robust Quishing Detection on Low-Resolution QR Codes via Feature Learning and Residual CNNs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Phuc Hao Do,
  • Huu Phu Le,
  • Vo Hoang Long Nguyen,
  • Nang Hung Van Nguyen

摘要

QR code-based phishing (“quishing”) is a growing cybersecurity threat. The baseline XGBoost model by Trad and Chehab (2025) [1] achieves a ROC-AUC of 0.9133 on 69x69 images. We propose a feature learning pipeline integrating image-based, metadata, and optional URL-based features, evaluated with Vanilla CNN, Residual CNN, and XGBoost models. Targeting a ROC-AUC above the baseline, our Residual CNN achieves 0.9313, followed by XGBoost (0.9158) and Vanilla CNN (0.8900), with superior robustness to perturbations like blur and compression. Contributions include an optimized feature pipeline for low-resolution QR codes and a comparison of deep learning and boosting models for real-world quishing detection.