Cybersecurity Framework using Deep learning to Detection of Malicious 2D Barcode Attacks

Abstract

Two-dimensional (2D) barcodes are being used in the development of many modern digital services, such as payment systems, authentication systems, and information access systems. However, the use of the barcode has been exploited by attackers, who insert malicious payloads in the barcode while keeping the structure of the barcode unchanged. This article proposes the application of deep learning model for detecting visually manipulated barcodes directly from the image data. Two models, namely a Convolutional Neural Network (CNN) model and a hybrid CNN and Long Short-Term Memory (LSTM) model, are considered for the purpose of detecting visually manipulated barcodes. A dataset of 80,160 images of barcodes is created, which includes 16,032 benign barcodes and four types of attack scenarios: encoded manipulation, payload injection, structural rotation, and adversarial noise. The results obtained from the experiment show that the CNN model has achieved an accuracy of 99.97%, and the CNN-LSTM model has achieved an accuracy of 99.98%.