Register-transfer-level Hardware Trojan classification boosted with gate-level features

Abstract

Hardware Trojan (HT) is an alarming hardware security threat which has gained increased awareness over the last decade. Due to the emerging threat of HT, ensuring trustworthiness in an integrated circuit (IC) has become an important aspect to be considered during manufacturing. Hence, the design process of ICs must be reviewed to avoid HT insertion by malicious third-party vendor. The purpose of this research is to develop a HT detection method in register-transfer-level (RTL) description with an improved HT coverage compared to the other previously proposed methods. The proposed method discovered HT branching statement in the RTL description by utilising a supervised machine learning classifier based on ten (10) proposed two-abstraction-level features. The proposed two-abstraction-level features relevant to HT characteristics included branching probability features extracted at RTL and net testability features extracted at gate-level (GL). The effectiveness of the proposed features in detecting HTs with 19 Trust-Hub benchmark circuits were demonstrated. The Minimum Redundancy Maximum Relevance (mRMR) feature selection algorithm was utilised to prove that the combination of the proposed features can achieve maximum accuracy (ACC) of 99.97% in detecting HTs during classifier training. To avoid overfitting issue, the trained classifiers were further evaluated with a classifier testing experiment on unseen circuit. The unseen circuit was completely independent of the training data, and it consisted of 24 HT circuits derived from a genuine Keccak encryption circuit. By using a set of proposed HT stealthiness assessment measures, the HT coverage of the classifiers was evaluated. The decision tree (DT) classifier with the two-abstraction-level features achieved the highest 87.5% HT coverage with 81.25% true positive rate (TPR), 88.44% true negative rate (TNR), and 88.24% ACC respectively. The results proved that the two-abstraction-level features outperformed single-abstraction-level features with higher HT detection coverage.