Multimodal Deep Learning Fusion for Accurate and Explainable Malware Family Classification

Identifying malware families is vital for predicting attack campaigns and creating effective defense strategies. Traditional signature-based methods are insufficient against new and evasive malware, highlighting the need for adaptive, multimodal solutions. This paper proposes a deep learning framework that fuses visual and static features through a ConvNeXt-Tiny backbone with cross-attention integration and incorporates calibration strategies such as snapshot ensembling, test-time augmentation, and per-class bias adjustment. The model is evaluated on two publicly available datasets: Malimg and Fusion Malware. The results demonstrate an outstanding accuracy of 99.69% on Malimg and 98.67% on Fusion, with macro F1 scores of 99.22% and 98.12%, respectively. Bias calibration improved the detection of difficult families on Malimg, and error analysis of Fusion identified challenges with polymorphic and underrepresented families. Overall, combining multimodal fusion with lightweight calibration enhances robustness and interpretability for real-world malware detection and attribution.