A HYBRID DEEP LEARNING APPROACH WITH SPATIAL ATTENTION MECHANISM FOR VISUAL-BASED MALWARE DETECTION IN IOT SECURITY

AYTURAN K.

Black Sea Journal of Engineering and Science, cilt.9, sa.3, ss.1256-1268, 2026 (TRDizin)

Özet

The rapid proliferation of embedded technologies and the Internet of Things (IoT) has triggered a dramatic escalation in the frequency and complexity of malicious activities targeting such networks. Conventional detection techniques relying on signatures often fall short when countering obfuscation and packing strategies commonly employed by cybercriminals. I present a novel transfer learning-driven neural network architecture which examines malicious software through the conversion of raw binary codes to generate grayscale visual representations. The original value of the study lies in the Spatial Attention mechanism integrated into the standard ResNet-50 architecture, which enables the network to prioritize salient texture patterns of malicious code. To validate the technique and its generalization capability, comprehensive experiments were conducted utilizing both the grayscale Malimg benchmark archive (9,339 samples, 25 families) and the RGB-based MaleVis dataset (26 families). Empirical findings demonstrate that the proposed hybrid framework yielded exceptional classification accuracies of 99.36% on Malimg and 97.25% on MaleVis, proving its robust dataset-independence. The findings reveal that visual analysis methods enhanced with attention mechanisms demonstrate higher performance than standard convolutional neural networks and offer an effective solution for detecting next-generation cyber threats.