Article
Identifying Internet of Things Software Activities using Deep Learning-based Electromagnetic Side-Channel Analysis
Contribution Summary
This paper investigates the potential of machine learning (ML) techniques in identifying complex activities on Internet of Things (IoT) devices using electromagnetic side-channel analysis (EM-SCA). The researchers generated a dataset by running ten different sorting algorithms on an Arduino Leonardo device, which represents a low-powered IoT device. The algorithms were continually sorting arrays of 100 elements randomly generated in ascending order. The dataset was used to train various classification models, including deep learning models, to predict the activity being executed from the electromagnetic noise emitted by the device. The results show that convolutional neural networks (CNNs) can accurately predict the activity being executed with a high level of accuracy (99.6%). Additionally, the study found that Random Forests and Deep Learning are suitable ML models for making predictions with EM-SCA. The research contributes to the field of digital forensics by providing a novel approach to acquiring forensic insights from IoT devices, which is essential for law enforcement and digital investigators.
Keywords: Electromagnetic side-channel analysis; Internet of Things; Deep learning; Random forests; Machine learning; Digital forensics; IoT forensics; EM-SCA
Abstract
Internet of Things (IoT) is becoming the new frontier in digital forensics due to the abundance of IoT devices appearing in day-to-day life. The diversity and complexity of IoT ecosystems pose a considerable challenge to digital investigators that demand novel approaches. Electromagnetic side-channel analysis (EM-SCA) has been proposed as a promising window to gather forensically useful information from IoT devices. Machine Learning (ML) techniques are instrumental when performing EM-SCA on IoT devices. Our work aims to investigate how machine learning can be applied to accurately identify complex activities on IoT devices from their generated electromagnetic noises. To this end, a range of classification models were created, including deep learning models, to predict the activity from the electromagnetic noise emitted while the device performed the activities. A dataset was generated by using ten different well-known sorting algorithms with diverse computational time complexities and running them on an Arduino Leonardo device to represent a low-powered IoT device. The algorithms were continually sorting arrays of 100 elements randomly generated in ascending order. Experiments were conducted to identify which ML methods performed better with the generated data sets. Furthermore, more experiments were conducted to identify how the methods perform depending on the window size of raw samples and the number of examples against which they are trained. From the experimental results, it is possible to predict which activity is being executed with a high level of accuracy (99.6%) with a convolutional neural network (CNN). It was also found that Random Forests (RF) and Deep Learning (DL) are suitable ML models for making predictions with EM-SCA.
BibTeX
@article{Le2021IoTEMSCA,
author={Le, Quan and Miralles-Pechu{\'a}n, Luis and Sayakkara, Asanka and Le-Khac, Nhien-An and Scanlon, Mark},
title="{Identifying Internet of Things Software Activities using Deep Learning-based Electromagnetic Side-Channel Analysis}",
journal="{Forensic Science International: Digital Investigation}",
volume = {39},
number={1},
pages = {301308},
year = 2021,
month=12,
publisher={Elsevier},
abstract={Internet of Things (IoT) is becoming the new frontier in digital forensics due to the abundance of IoT devices appearing in day-to-day life. The diversity and complexity of IoT ecosystems pose a considerable challenge to digital investigators that demand novel approaches. Electromagnetic side-channel analysis (EM-SCA) has been proposed as a promising window to gather forensically useful information from IoT devices. Machine Learning (ML) techniques are instrumental when performing EM-SCA on IoT devices. Our work aims to investigate how machine learning can be applied to accurately identify complex activities on IoT devices from their generated electromagnetic noises. To this end, a range of classification models were created, including deep learning models, to predict the activity from the electromagnetic noise emitted while the device performed the activities. A dataset was generated by using ten different well-known sorting algorithms with diverse computational time complexities and running them on an Arduino Leonardo device to represent a low-powered IoT device. The algorithms were continually sorting arrays of 100 elements randomly generated in ascending order. Experiments were conducted to identify which ML methods performed better with the generated data sets. Furthermore, more experiments were conducted to identify how the methods perform depending on the window size of raw samples and the number of examples against which they are trained. From the experimental results, it is possible to predict which activity is being executed with a high level of accuracy (99.6%) with a convolutional neural network (CNN). It was also found that Random Forests (RF) and Deep Learning (DL) are suitable ML models for making predictions with EM-SCA. },
doi={10.1016/j.fsidi.2021.301308},
}