Enhanced Quishing Attack Detection using CNN
Year: 2026
Type: conference-paper
Abstract: The use of QR codes in finance and in business is ubiquitous, and as such that has made something known as Quishing an even bigger issue. It is just phishing only that it uses codes on the QR, not emails or links. Text-based phishing is detected by filters, whereas Quishing deceives since the code itself appears to be harmless, thus users simply move their eyes over it without even considering the security issue. I believe the problem is the way that the existing defenses operate. Others only check the URL after scanning and, perhaps, clicking, and that is why you are already compromised in case it is bad. Others apply machine learning to very small datasets that reduce the QR to the simplest and do not capture the actual details in the picture. This paper attempts to overcome that by adopting a computer vision solution which detects Quishing immediately before there is any scanning of the code i.e. at the first sight of the code on the computer screen. We also created a massive dataset, 600,000 images 224 by 224 pixels and drew it from real threats and regular harmless ones. This sounds like a lot, but it includes actual variety. It is a custom CNN that is being trained with all such data. It achieved 99.62 percent accuracy and Area Under the Curve (AUC) score of 0.9982 on a test set of about 126,000 images. Pretty high, right? It implies that the process of deep learning is capable of isolating the safe QR codes among the malicious ones without even reading them. In the case of mobile devices, it may be real time protection that is fast. I am not completely aware that it is managing all the edge cases yet. It is significant on their visual side, as text filters do not fit here.
Authors: Rudra Kumar, Shivangi Sehgal, Ravi Sharma
Institution: Galgotias University
Language: English
Topic: Spam and Phishing Detection
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7171519104
