A Novel Deep Ensemble Framework for IoT Malware Variant Detection
Hamza Javed, Feng Zeng, Muhammad Shaheer · 2024
Malware presents formidable challenges to digital security, underscoring the need for robust detection mechanisms to counter potential cyber threats. Relying solely on a single model for detection, however, comes with limitations in accurately capturing the complexity and variability of malware samples. In response, this study introduces a novel ensemble method for malware classification, harnessing the combined strengths of multiple deep learning models to bolster detection accuracy and resilience. Through meticulous fine-tuning of established pre-trained models like MobileNetV1, MobileNetV2, and Xception, we integrate their predictive capabilities using a stack ensemble framework. This ensemble approach aggregates decisions from the three models through a stacking technique, yielding outcomes characterized by high generalizability and minimal variance. Our methodology innovatively avoids the necessity for conventional techniques such as feature engineering and other domain-specific methods traditionally employed in malware classification. Our model's efficacy is rigorously assessed on the Malimg dataset, encompassing 9,335 images spanning 25 distinct classes. Demonstrating superior performance compared to individual models, our approach achieves an impressive accuracy rate of 98.35% on the Malimg dataset. Additionally, statistical analyses, including McNemar's test, substantiate the significance of our findings. Comparative evaluations against existing methods further underscore the efficacy of our proposed ensemble approach in effectively detecting malware instances.