AQRACE: MACHINE-LEARNING-BASED DETECTION OF MALICIOUS URLS IN QR CODES FOR REAL-TIME SECURITY ASSESSMENT

Abstract

The widespread use of Quick Response (QR) codes has increased exposure to QR-based phishing, malicious redirection, and malware distribution because users cannot readily inspect the destination before scanning. This study develops aQRace, a machine-learning-based QR code security application that analyses embedded Universal Resource Locators (URLs) before users access their destinations. The study combines URL-based feature extraction, comparative machine-learning evaluation, and operational system integration. URL characteristics covering structural, linguistic, technical, and advanced properties were extracted for classification. Four supervised machine-learning algorithms–XGBoost, Random Forest, Decision Tree, and Logistic Regression–were trained and evaluated using a balanced dataset of 14,000 URLs comprising 7,000 safe and 7,000 malicious samples. Model performance was compared using accuracy, precision, recall, and F1-score. XGBoost achieved the highest accuracy of 84% and precision of 0.80, whereas Random Forest achieved the highest recall of 0.95. Both models achieved an F1 Score of 0.84. XGBoost was selected for deployment based on its overall performance profile and integrated into a client-server application comprising a Flutter mobile frontend and Flask cloud backend. The application supports QR code image uploads and camera-based scanning, and performs URL processing, feature extraction, classification, and user warnings before destination access. Functional testing with representative safe and malicious QR codes confirmed the operation of the complete detection workflow. The findings demonstrate that conventional machine learning using URL-based features can support pre-access QR code security assessment while providing an operational alternative to increasingly complex deep learning approaches.