Download
Abstract
Detecting malware in Internet of Things (IoT) systems is challenging due to resource limitations. As a result, detection and classification models are often deployed in cloud environments or distributed across devices and servers. In distributed scenarios, transferring model weights can create bottlenecks, with packet loss risking training stability. This research presents a modified Visual Transformer (ViT) that utilizes Rotational Positional Encoding (RoPE), which takes into account network bandwidth, quality, and device status. Our model effectively resolves bottleneck issues while ensuring accurate malware detection. Under extreme network stress, our approach on a medium ViT achieves 82.1% classification accuracy, compared to 65.2% for the standard RoPE—a 16.9 percentage point improvement. In micro-scale architectures, Modified RoPE reaches 97.6% accuracy versus 84.0% for Standard RoPE, representing a 13.6 percentage point improvement. These results confirm that our regularization methodology effectively maintains spatial relationships during network degradation while improving positional encoding across various architectural scales.
Citation
NOT PUBLISHED YET