Federated and Explainable Quantum Machine Learning for Mobile Malware Detection
Abstract
Malware detection in mobile environment represents a crucial task, considering the inefficacy on signature-based antimalware to detect zero-day attacks. Researchers from academic and industrial community are proposing approaches based on machine learning, obtaining interesting performance in mobile malware detection but these approaches face several limitations in real-world mobile devices deployment i.e., privacy concerns arising from the need to centralize sensitive operational data and limited explainability that reduces trust from malware analyst and final user point of view. With the aim to address these challenges, in this paper we propose a privacy-preserving and explainable mobile malware detection method considering Federated Machine Learning by learning a Hybrid Quantum-Classical Neural Network developed by authors. The proposed method considers the transformation of Android applications into images (but the method can easily adapted to different operating systems) and locally processed at each client (i.e., mobile device), ensuring that user application never leaves its origin. The convolutional backbone extracts spatial features, while the variational quantum circuit enhances non-linear representational capacity within the federated setting. To provide explainability, we consider a multi-level explainability suite combining Class Activation Mapping, quantum-state analysis through Pauli-Z expectation values and per-qubit saliency mapping. Experimental results on a dataset composed by real-world Android malware and trusted application demonstrate that the approach is able to reach interesting detection performance, by obtaining an accuracy ranging from 0.72 to 0.81, thus showing the effectiveness of the proposed method. Furthermore, the explainability provides actionable insights into both classical and quantum feature contributions, facilitating trust and enabling practical deployment within real-world mobile malware detection.