Imbalanced Hardware Trojan Detection Based on Conditional Generative Adversarial Networks

Abstract

: Hardware Trojan (HT) can compromise the security of a system by changing the integrated circuit (IC) functionality and reducing the system ꞌ s reliability. To handle this issue, machine learning has been widely used to analyze the datasets extracted from circuits to detect hardware Trojans. However, the extant HT detection methods provide low performance and are not applied to evaluate comprehensively using imbalanced data, which may degrade the performance of machine learning. To overcome this limitation, we proposed a conditional generative adversarial networks method that integrates the machine learning with the deep learning to detect the hardware Trojans injected in Register-Transfer Level code. A framework including feature extraction and data augmentation is proposed. Firstly, the control flow graph and data dependence graph are constructed from Register-Transfer Level code. Then, the 16 features are extracted by walking the graphs. Because there is class imbalance, a Conditional Generative Adversarial Network is proposed. Again, based on the Conditional Generative Adversarial Network model, the synthetic data is generated to balance the feature datasets. Furthermore, machine learning algorithms analyze the balanced feature datasets. The experiments use the Trust-hub benchmarks and Hummingbird e203 designs to assess our method. Finally, compared to the original datasets, the datasets enhanced by our proposed CGAN improved the F1 score and GMean of the machine learning algorithms by 32.31% and 24.17%, respectively. Moreover, when compared to the SMOTE-enhanced datasets, our method yielded a 30.51% increase in F1 score and a 21.98% increase in GMean. This demonstrates the consistency and effectiveness of our newly proposed model in detecting different types of HTs across imbalanced dataset, and it contributes to enhancing the security and trustworthiness of ICs against hardware Trojan attacks