Deep Learning-Based Detection and Classification of Fake QR Codes

Abstract— The objective of Fake QR Code Detection through Deep Learning is to accurately identify and classify harmful QR codes that pose a threat to user security by employing advanced computer vision methodologies. To achieve optimal reliability and effectiveness, the initiative implements a multi-tiered deep learning framework utilizing a carefully selected Kaggle dataset comprising both authentic and counterfeit QR codes. The process initiates with the YOLOv8 model for QR code detection, which isolates QR code regions within input images. Subsequently, these detected QR codes are processed by a specialized classification network built on the DenseNet121 architecture, which has been pre- trained on ImageNet and specifically fine-tuned to distinguish between legitimate and fraudulent QR codes. The classification model demonstrates a commendable accuracy rate of 92%, indicating strong generalization capabilities. Additionally, a phishing URL detection mechanism is integrated to verify the destination links embedded in authentic QR codes, thereby further bolstering security. This comprehensive approach not only illustrates how deep learning can mitigate the risks associated with the proliferation of counterfeit QR codes but also highlights its potential applications in digital marketing, insecure mobile transactions, and access control systems. Keywords: Deep Learning, DenseNet121, YOLOv8, CNN, URL phishing detection