From One Network to Another: Transfer Learning for IoT Malware Detection

The rapid proliferation of IoT technology expands the attack surface, increasing the number and type of vulnerabilities that malicious actors can exploit. Machine Learning and Deep Learning approaches can detect IoT malware, however they often require extensive data collection from the target IoT network—a time-consuming, resource-intensive process that also introduces security risks. This work explores Transfer Learning to leverage knowledge from publicly available IoT network datasets for securing previously unseen networks. Our results show that infusing just a few samples from the target network can boost detection accuracy by up to $40 \%$, reaching $95 \%$ accuracy. Notably, significant improvements occur with as few as 6 samples per class (12 for binary malware detection), while 1,000 samples per class achieve peak performance independent of the training dataset size. Based on our findings, we strongly propose Few-Shot Transfer Learning as an effective and practical approach for IoT malware detection.