SC-COTD: Hardware Trojan Detection Based on Sequential/Combinational Testability Features using Ensemble Classifier

Abstract

Security against Hardware Trojans (HT) is an important concern in integrated circuits (IC) design and fabrication. Most of the current HT detection methods are based on the golden model of circuit design. Further, some approaches require test pattern for HTs activation. In this paper, we propose SC-COTD (Sequential/Combinational Controllability and Observability features for hardware Trojan Detection), an effective hardware Trojan detection to get rid of both golden chip and test pattern limitations. SC-COTD uses both sequential and combinational testability measures to detect and locate HT signals by a machine learning approach. This method deploys an ensemble classifier based on k-means clustering. The clustering models have diverse variety in testability features along with size of clustering which inspect and reveal different aspects of netlist conventional for a collaborative scheme. The clustering results are filtered and then fed into a decision-making procedure based on majority voting to eliminate the limited flaws of each model. The evaluation results on TrustHUB benchmarks demonstrate that, SC-COTD can detect and locate HTs with 100% without any false negative, i.e., Recall = 1. Although our method has a limited number of false positive, it has the best performance in comparison to well-known previous approaches.

Data Augmented Hardware Trojan Detection Using Label Spreading Algorithm Based Transductive Learning for Edge Computing-Assisted IoT Devices

“…The improved TPR values can be attributed to the utilization of initial cluster infor- Semi-supervised learning TPR(%) [37] TPR(%) [38] TPR(%) [46] TPR(%) [31] TPR(%) [30] TPR pares the performance of the proposed work with existing supervised schemes such as [37], [38], unsupervised schemes [31], [30] and few-shot learning based schemes [46] in terms of TPR. The method outperforms [37], [31] and [30] by 4.03% , 16.62% and 11.9% in terms of TPR. The method achieves comparable performance in comparison with [38].…”

“…Testability based HTD approaches was further extended to incorporate structural features [28] or fault modelling techniques [29] to handle the aforementioned limitations. Another referencefree HT detection scheme utilizing testability measures was developed in [30]. Further, information theory-based HT detection approach investigating the relation between transition probability and the information available on a net for unsupervised Trojan detection using density-based clustering algorithm was attempted in [31].…”

“…In addition, there is no unified method of labeling the nets, leading to discrepancies in result interpretation. Unsupervised strategies, in general, adopt testability measurebased features targeting Trojans that have low controllability and low observability [30]. Such methods can be circumvented by redesigning the Trojans to satisfy the conditions of a normal circuit, as mentioned in [21].…”

“…It does not consider the impact a feature has on the model's predictive capability. For measuring the predictive capability of the model, permutation importance [30] is adopted. It calculates model dependency on the features separately.…”

“…The valuable prior information in the labeled data has been exploited in the proposed semi-supervised algorithm to enhance the TPR when compared to [43]. The improved TPR values can be attributed to the utilization of initial cluster infor- Semi-supervised learning TPR(%) [37] TPR(%) [38] TPR(%) [46] TPR(%) [31] TPR(%) [30] TPR pares the performance of the proposed work with existing supervised schemes such as [37], [38], unsupervised schemes [31], [30] and few-shot learning based schemes [46] in terms of TPR. The method outperforms [37], [31] and [30] by 4.03% , 16.62% and 11.9% in terms of TPR.…”