MalSFF: Multi-architecture malware detection using multi-static feature fusion based on image visualization and learning methods

Malware detection is a necessity in the modern digital world. This research presents a novel MalSFF: Multi-Architecture Mal ware Detection Using Multi- S tatic F eature F usion Based on visual image analysis and transfer learning. Firstly, it decompiles binary programs to extract bytecodes and assembly code (ASM) through reverse-engineering before transforming them into grayscale images. This research strategically fine-tunes the MobileNet models (V1, V2, V3-Small, and V3-Large) for feature extraction of both file types. Thereafter, it performs feature stacking through early fusion, late fusion, and ensemble voting to obtain a single feature map, and then utilizes a filter-based feature selection algorithm. Finally, the MalSFF employs six different classifiers, with optimized hyperparameters using an automated grid-search algorithm. For better generalization, this study uses four different datasets: (i) Microsoft BIG, (ii) MalImg, (iii) Dumpware10, and (iv) Real-world samples. The MalSFF achieved 98.72% accuracy for the MalImg and 96.93% accuracy, 97% precision, 97% recall, and 97% F1-score, 0.012 ms of response time for the BIG dataset. For memory-resident malware, it achieved 93.84% accuracy and a 91% F1-score, with a response time of only 0.05 s. The MalSFF demonstrates resilience against FGSM, PGD, and DeepFool adversarial attacks. The MalSFF is a lightweight and computationally efficient, well-suited for resource-constrained IIoT networks.