Hardware Trojan Detection Based on Collaborative Optimization of Feature Fusion and Sample Screening

Globalization exacerbates intellectual property (IP) piracy and hardware Trojan (HT) insertion in circuit design. Existing machine learning methods for HT detection can spot malicious insertions but often overlook the issue of generating low-quality synthetic samples in the face of sample imbalance. Moreover, using a single detection model limits the ability to detect a variety of HT. In this paper, we first propose six new HT features. Then, to address the sample imbalance issue and generate higher-quality synthetic samples, we employ the borderline synthetic minority over-sampling technique (Borderline-SMOTE) to generate synthetic samples, selecting the top 90% of synthetic samples for in-depth analysis. Furthermore, this study uses neural network technology for the efficient fusion of sample features, forming an advanced feature containing more comprehensive information, which is then applied as another new feature in our detection model. Finally, we use the light gradient boosting machine (LightGBM) algorithm for the ultimate HT detection. Through a series of Trust-HUB benchmark tests, our method has demonstrated outstanding performance, with the average true positive rate (TPR) of 92.83%, achieving 100% in several circuit tests. Additionally, the average true negative rate (TNR) is 90.26%, with the average accuracy of 90.86%.