💻 computer science

Quantum Machine Learning in Cybersecurity: A Hybrid Convolutional Neural Network with Skip Connections for QR Code Phishing Detection

This study proposes a Hybrid Quantum-Classical Convolutional Neural Network with skip connections (HybridSkip) that directly classifies QR code images based on visual structure to detect phishing, achieving significantly higher recall and F1-scores with fewer parameters than classical baselines, thereby offering a resource-efficient solution for proactive quishing defense.

Original authors: Md Farhan Shahriyar

Published 2026-09-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Md Farhan Shahriyar

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern digital world, a simple square of black and white squares has become a universal key. These are Quick Response codes, or QR codes, the two-dimensional barcodes found on restaurant menus, payment terminals, and transit tickets. They are designed to be scanned instantly by a smartphone camera, bridging the physical and digital worlds without a single keystroke. However, this convenience has created a new vulnerability. Because the human eye cannot read the data hidden inside the pattern, a malicious actor can replace a legitimate code with a fake one that leads a victim to a fraudulent website. This specific type of attack, known as "quishing," bypasses traditional email filters and link scanners because the threat arrives as a printed image, not a digital message. The core challenge for security experts is that by the time a user scans the code, the damage is often already done, as the phone automatically opens the hidden link.

Researchers have long tried to solve this by decoding the image to read the hidden URL and then checking if that address is dangerous. But this method fails if the code is damaged, if the internet is unavailable, or if the bad link was created just moments before the scan. A new approach suggests looking at the picture itself rather than the text inside it. The idea is that the length and complexity of the hidden message leave a subtle statistical fingerprint on the arrangement of the black and white squares. A computer trained to recognize these visual patterns could spot a malicious code before it is ever decoded. This is where the intersection of cybersecurity and quantum computing enters the story. Quantum machine learning uses the principles of quantum mechanics to process information in ways that classical computers cannot, potentially offering a more efficient way to find patterns in small, complex datasets.

Md Farhan Shahriyar, a researcher at Hajee Mohammad Danesh Science and Technology University in Bangladesh, has developed a new system that combines these ideas to detect quishing attacks. The system is a hybrid model, meaning it mixes a standard classical computer program with a specialized quantum component. The classical part acts as a set of eyes, scanning the image of the QR code to identify its basic shapes and structures. The quantum part then takes a compressed version of those shapes and processes them through a simulated quantum circuit, a mathematical structure that can explore many possibilities at once. The goal was to see if this combination could identify a fake code more accurately than a standard computer program, while using fewer computing resources.

The study tested this system on a dataset of one thousand QR code images, split evenly between safe codes and malicious ones. The researchers trained the model to look at the raw images without ever trying to read the text inside them. They compared their new hybrid design against several other models, including standard computer-only programs and other hybrid versions without a specific architectural tweak. The most successful version of their system, which they named HybridSkip, introduced a "skip connection." In simple terms, this is a shortcut that allows the initial visual information from the classical part of the system to bypass the quantum processor and go straight to the final decision-making stage. This design ensures that no useful detail is lost during the complex quantum processing and helps the system learn more effectively.

The results showed that this new approach was highly effective. The HybridSkip model correctly identified malicious QR codes with a higher success rate than the best classical computer-only models. Specifically, it caught about 72 percent of the bad codes, compared to roughly 65 percent for the strongest traditional model. This improvement is significant in a security context because missing a bad code means a user could be tricked into giving away their passwords or money, while a false alarm only costs a few extra seconds of checking. Furthermore, the new model achieved this higher accuracy while using nearly 38 percent fewer adjustable settings, or parameters, than the classical models. In the world of artificial intelligence, having fewer parameters usually means the system is faster and requires less memory, making it easier to run on mobile phones or small devices.

The researchers confirmed that these improvements were not just a lucky fluke by running the tests multiple times with different splits of the data. Statistical analysis showed that the gains in detecting bad codes were real and consistent. The study also noted that the system works entirely offline. Because it analyzes the image directly, it does not need an internet connection to check a database of known bad links. This makes it particularly useful for scenarios like scanning a code at a remote kiosk or a point-of-sale terminal where a network connection might be unreliable.

While the results are promising, the study acknowledges certain limitations. The quantum part of the system was simulated on a standard computer, not run on actual quantum hardware, which is still an emerging technology. The researchers also noted that the dataset used was relatively small, containing only one thousand images, and that real-world codes can vary widely in quality, lighting, and damage. Future work will need to test the system on larger collections of images and on actual quantum devices to see how it holds up under real-world conditions. Nevertheless, the study demonstrates that a hybrid approach, which carefully blends classical and quantum processing, can offer a more efficient and effective way to spot visual threats. By treating the QR code as a picture rather than a message, this new method provides a proactive layer of defense that could help keep users safe in an increasingly connected world.

Technical Summary: Hybrid Quantum-Classical CNN for QR Code Phishing Detection

Problem Statement
QR code phishing ("quishing") has emerged as a critical cybersecurity threat, exploiting the opacity of QR codes to deliver malicious payloads without triggering traditional email or link-based defenses. Current countermeasures are largely reactive and flawed: they rely on decoding the QR code to extract a URL for blacklist checking or reputation analysis. This approach fails when the destination is a zero-day domain, when the URL is obfuscated, or when the scanning device lacks network connectivity (e.g., offline kiosks). Furthermore, structural metadata-based detection fails if the code is physically damaged or deliberately obfuscated. While image-based classification offers a decode-free alternative, purely classical deep learning models often overfit on the small, labeled datasets typical of security research.

Methodology
The study proposes a Hybrid Quantum-Classical Convolutional Neural Network (HQCNN) that classifies QR codes directly from their visual pixel data, bypassing the decoding step entirely. The architecture consists of three main components:

Classical Convolutional Backbone: A five-layer CNN processes grayscale QR images (48×48 pixels) to extract hierarchical spatial features. These features are projected via a fully connected layer and a tanh activation function to constrain values to the range $[-1, 1]$ , suitable for quantum angle embedding.
Variational Quantum Circuit (VQC): The projected classical features serve as input to a parameterized quantum circuit. The circuit uses angle embedding (encoding features as rotation angles of $RY$ gates) followed by $L$ variational layers. Each layer consists of trainable $RY$ and $RZ$ rotations on all qubits, coupled with a CNOT entangling ladder. The circuit outputs Pauli-Z expectation values, which are fed into a classical classification head.
HybridSkip Architecture: A novel variant introduces a residual bypass (skip connection). In this design, the classical backbone features ( $z$ ) are routed directly to the classification head in parallel with the quantum-processed path ( $q$ ). The two streams are concatenated ( $h = [q; z]$ ) before the final classification layer. This design aims to preserve gradient flow (mitigating the "barren plateau" risk), retain spatial context lost during low-dimensional quantum encoding, and eliminate the need for a wide compensatory fully connected layer.

The models were evaluated on a balanced corpus of 1,000 QR code images (500 benign, 500 malicious) using stratified five-fold cross-validation and a held-out test set. Six configurations were compared: two classical baselines, two standard HQCNNs, and two HybridSkip variants, varying in qubit count (4 vs. 6) and variational layers (3 vs. 4).

Key Contributions

First Decode-Free Quantum Visual Classifier: This work presents the first hybrid quantum-classical architecture for QR code phishing detection that operates solely on pixel data, requiring no decoding or network connectivity.
HybridSkip Design: The introduction of a residual bypass in hybrid quantum-classical networks, which improves gradient flow and parameter efficiency by allowing the classifier to access raw classical features alongside quantum features.
Parameter-Efficient Benchmarking: A rigorous, parameter-normalized comparison demonstrating that the HybridSkip architecture achieves superior performance with significantly fewer trainable parameters than classical baselines.
Statistical Rigor: The study employs paired statistical tests (t-tests, Wilcoxon signed-rank), effect size analysis (Cohen's d), and confidence intervals across cross-validation folds, moving beyond simple point estimates to validate improvements on small datasets.

Results
The HybridSkip-6q4L configuration (6 qubits, 4 variational layers) demonstrated the most favorable trade-off between accuracy and model size:

Performance: It achieved a cross-validation F1-score of 0.694 ± 0.027 and a recall of 0.710 ± 0.031. On the held-out test set, it reached an F1-score of 0.715 and recall of 0.724.
Comparison: Compared to the strongest classical baseline (Classical-h8), HybridSkip-6q4L improved recall by 6.8% and F1-score by 3.3% (in cross-validation).
Efficiency: The model utilized 21,015 trainable parameters, a 37.6% reduction compared to the classical baseline (33,673 parameters).
Significance: Statistical analysis confirmed that the recall gains were significant (p < 0.05) with large effect sizes (Cohen's d > 0.8). The model occupies the Pareto front for accuracy versus parameter count.

Significance and Claims
The paper claims that parameter-efficient hybrid architectures offer a viable path for proactive quishing defense on resource-constrained endpoints (e.g., mobile devices, embedded readers) in the Noisy Intermediate-Scale Quantum (NISQ) era.

Key claims include:

Offline Capability: The model functions without network connectivity, addressing a critical failure mode of URL-based defenses.
Gradient Preservation: The skip connection successfully mitigates gradient attenuation in the quantum bottleneck, a common issue in hybrid models.
Generalization: The quantum feature map allows the model to generalize from limited data (1,000 samples) more effectively than classical models of similar capacity, likely due to the implicit high-dimensional feature mapping of the Hilbert space.
Deployment Readiness: While the quantum component is currently simulated, the parameter savings are realizable on classical hardware today, and the architecture is designed to be compatible with future physical quantum devices.

The author maintains modesty regarding absolute performance, acknowledging that F1-scores (0.63–0.72) are lower than URL-based detectors because visual classification is a strictly harder problem that lacks the payload string. They emphasize that the primary contribution is the improvement in recall and efficiency relative to classical image-based baselines, rather than surpassing URL-based methods. Future work is identified as needing larger, diverse datasets, noise-aware training for real hardware, and adversarial robustness testing.

Technical Summary: Hybrid Quantum-Classical CNN for QR Code Phishing Detection

More like this