Search papers, labs, and topics across Lattice.
This research introduces a scalable Quantum Kernel-based Machine Learning (QKML) framework designed for effective malware family classification, addressing the challenges posed by the increasing structural similarity and obfuscation of malware samples. By employing a supervised Linear Discriminant Analysis (LDA) for feature extraction and a fidelity-based quantum kernel for capturing nonlinear relationships, the framework achieves an impressive classification accuracy of 80.88%. The use of the Nystr枚m approximation method allows for efficient multiclass classification without the computational burden typically associated with kernel matrix construction, significantly improving performance over classical machine learning approaches.
Scalable quantum kernel methods can dramatically enhance malware classification accuracy, achieving over 80% performance while sidestepping traditional computational bottlenecks.
The classification of malware families is a key challenge in cybersecurity, which enables threat attribution, analysis of attack operations, and the formulation of effective defense strategies. Emerging malware samples are becoming increasingly structurally similar and obfuscated, making accurate multiclass classification challenging for traditional machine learning models, especially when deployed at scale. In this research, we propose a scalable Quantum Kernel-based Machine Learning (QKML) framework for malware family classification that addresses both accuracy and efficiency constraints. The proposed framework extracts structural features from executable files and uses a supervised Linear Discriminant Analysis (LDA) projection to generate a compact, class-aware representation well suited for quantum processing. The nonlinear relationships among malware families are captured using a fidelity-based quantum kernel built from parameterized quantum circuits. We use the Nystr枚m approximation method to obtain a low-rank approximation of the quantum kernel, which enables effective multiclass classification via ridge regression and enables learning from all available training samples without incurring the quadratic computational cost of kernel matrix construction. The proposed model achieves strong classification performance, with 80.88% accuracy, outperforming classical machine learning baselines under identical feature and data splits, according to experimental evaluation on a large-scale malware dataset that includes 18,836 samples across 23 malware families. These findings suggest that scalable quantum-kernel-based machine learning can offer measurable performance advantages for real-world malware family classification tasks.
