Visual Authenticity Detection of QRIS Codes Using Convolutional Neural Networks

Abstract

The Quick Response Code Indonesian Standard (QRIS) has rapidly expanded as Indonesia’s national digital payment system, with more than 30 million merchants registered by 2024 to promote financial inclusion and cashless transactions. However, increasing fraud incidents—through both digital manipulation and physical tampering such as counterfeit stickers—have exposed vulnerabilities in user verification and limited digital literacy among Micro, Small, and Medium Enterprises (MSMEs). This study proposes a Convolutional Neural Network (CNN)–based deep learning system for detecting the visual authenticity of QRIS codes. The dataset comprises 60 images, including 30 genuine, 17 dummy, and 13 tampered QRIS codes. Preprocessing involved grayscale normalization, resizing to 128×128 pixels, and noise filtering. The CNN, adapted from MobileNetV2, was trained for 50 epochs using the Adam optimizer with a 0.001 learning rate. Experimental results show a training accuracy of 98.7% and testing accuracy of 94.3%, with a hybrid verification accuracy of 95% when combined with EMVCo-based payload validation. The system correctly identifies QRIS images with clear structures and valid patterns, even if they are dummy or non-official codes. Conversely, genuine merchant QRIS may be flagged as invalid if not listed in the application’s whitelist. Due to restricted access to Bank Indonesia’s official merchant database, this research is limited to visual and structural validation rather than full merchant authenticity verification. Nonetheless, the proposed approach provides a practical and lightweight tool for preliminary QRIS verification and strengthens user awareness of visual payment security.