Resumen
The growth of the Internet of Things (IoT) has led to a significant rise in cyber attacks and an expanded attack surface for the average consumer. In order to protect consumers and infrastructure, research into detecting malicious IoT activity must be of the highest priority. Security research in this area has two key issues: the lack of datasets for training artificial intelligence (AI)-based intrusion detection models and the fact that most existing datasets concentrate only on one type of network traffic. Thus, this study introduces Dragon_Pi, an intrusion detection dataset designed for IoT devices based on side-channel power consumption data. Dragon_Pi comprises a collection of normal and under-attack power consumption traces from separate testbeds featuring a DragonBoard 410c and a Raspberry Pi. Dragon_Slice is trained on this dataset; it is an unsupervised convolutional autoencoder (CAE) trained exclusively on held-out normal slices from Dragon_Pi for anomaly detection. The Dragon_Slice network has two iterations in this study. The original achieves 0.78 AUC without post-processing and 0.876 AUC with post-processing. A second iteration of Dragon_Slice, utilising dropout to further impede the CAE?s ability to reconstruct anomalies, outperforms the original network with a raw AUC of 0.764 and a post-processed AUC of 0.89.