Abstract
The number of malware attacks on IoT devices has increased significantly in recent years. However, it is difficult to implement recent anti-malware software on these devices due to limitations in computer resources. Furthermore, the number of samples of each malware family in a dataset is often imbalanced. This study measured the improvement in accuracy achieved by implementing a machine learning-based malware detection system in C/C++ for IoT devices after correcting for an unbalanced sample count in the dataset using information gained from byte n-grams. Our method was implemented on real IoT devices and the resource consumption was evaluated. The SVM model with the best performance achieved an F1-score of 100 % for malware detection. Inference time was 5,108 seconds and memory usage was 103.9 MB when an SVM model with a Top-L of 50,000 was installed on a Raspberry Pi Zero 2 W. On the NanoPi R2S-LTS, the inference time was 4,072 seconds, with a memory usage of .