Optimized Attention Assisted Network for Forensic-Aware Malware Detection in Consumer Electronics

Download PDF
Open Webpage

Hazrat Bilal, Sushil Kumar Singh, Abdullah M. Alnajim, Muhammad Islam, Ahlam Almusharraf, Lucia Cascone, Faisal Mehmood

Published: 01 Jan 2025, Last Modified: 09 Nov 2025IEEE Transactions on Consumer ElectronicsEveryoneRevisionsCC BY-SA 4.0
Abstract: The rapid evolution of malware, driven by advanced evasion techniques such as polymorphism, metamorphism, and adversarial attacks, poses a significant challenge to traditional malware detection (MD) methods. While deep learning (DL)-based approaches have demonstrated promising results, their reliance on suboptimal feature representations and computationally expensive attention mechanisms limits their adaptability and feasibility in real-world applications, particularly for resource-constrained consumer electronics. To address these limitations, we propose a novel MD framework that enhances efficiency and robustness through optimized feature extraction and attention mechanisms. First, we conduct an extensive empirical analysis to determine the most effective lightweight feature representation, selecting EfficientNetB3 as an optimal balance between performance and computational cost. Second, we introduce the Optimized Efficient Channel Attention (ECA) mechanism, which features a reduced kernel size, effectively capturing essential channel-wise dependencies while minimizing computational overhead. Third, we design a Spatial Attention (SA) module to refine malware representations, coupled with a Dilated Convolution Block (DCB) to expand the receptive field without excessive parameter growth, ensuring improved detection of malware patterns across multiple scales. We further boost detection performance by introducing a data augmentation pipeline that enhances model generalization and robustness against diverse malware variants. These techniques include rotation transformations, horizontal/vertical flipping, random cropping, contrast modification, and Gaussian noise injections. Finally, we validate our approach through extensive experiments on benchmark malware datasets, demonstrating superior classification performance and real-time feasibility for consumer electronics security. The proposed framework provides a scalable, adaptive, and privacy-preserving MD solution, offering enhanced resilience against emerging cyber threats in connected devices.