XAI-guided quantum-classical fusion for malware detection: integrating QSVM with transformer architectures

Abstract

The proliferation of sophisticated malware variants necessitates detection frameworks that balance accuracy with interpretability. This paper proposes a hybrid framework integrating global whale optimization, quantum support vector machine (QSVM), and transformer architectures with explainable artificial intelligence (XAI) components. The approach addresses feature selection, quantum-enhanced classification, and sequential pattern recognition through a unified optimization strategy. Evaluated on EMBER 2018 (1.1M samples) and UGRansome (207 K samples), the framework achieves 99.95% and 99.99% accuracy respectively, with ablation studies confirming synergistic performance gains. Statistical validation through 5-fold stratified cross-validation, 95% bootstrap confidence intervals, and McNemar's significance tests confirms that improvements over every baseline are statistically robust (p<0.001). The integrated XAI framework provides multi-level explanations through SHapley Additive exPlanations (SHAP) analysis and attention visualization, offering actionable insights for security analysts without compromising detection performance. This research contributes to sustainable development goals (SDG) 16 by promoting secure and resilient infrastructures in digital systems and enhancing institutional capacity to combat evolving cyber threats.