Lightweight Machine Learning Techniques forHardware Trojan Detection in IoT Devices

Abstract

Hardware trojans are malicious pieces of software
that attempt to prevent the normal operation of a chip, and are
carefully engineered not to be detected during the silicon design
and verification phase before it is actually sold to a consumer.
The military, business, and academics are all looking into this
new threat. Consequently, as a defense during chip deployment,
run-time hardware Trojan identification is vitally needed. This
work focuses on hardware Trojans that affect processor
performance. This study presents a machine learning-based
approach to detecting hardware Trojans in IoT devices by
exploiting the Hardware Trojan Dataset. To preserve the most
relevant features, the dataset was subjected to a thorough
preparation procedure that included data cleaning,
augmentation, label encoding, normalization, and feature
selection using PCA. A number of models were evaluated,
including Logistic Regression, ResNet, Decision Tree, and Long
Short-Term Memory (LSTM). With the highest accuracy,
precision, recall, and F1-score of 95.25%, 95.25%, 95.27%, and
95.25%, respectively, the LSTM model fared better than the
others. The outcomes demonstrate how well feature selection
and sequential deep learning architectures work together to
capture temporal relationships in power trace data. Overall, the
suggested approach shows a strong and trustworthy foundation
for improving IoT hardware security against Trojan assaults.