FAST-GO: Fast, Accurate, and Scalable Hardware Trojan Detection using Graph Convolutional Networks

Ali Imangholi1, Mona Hashemi2, Amirabbas Momeni1, Siamak Mohammadi3, Trevor E. Carlson4
1School of ECE, College of Eng., University of Tehran, 2School of ECE, College of Eng., University of Tehran, and School of Computing, National University of Singapore, 3School of ECE, College of Eng., University of Tehran, and School of Computing Science, IPM, 4National University of Singapore


The continuous advancements in technology and consequently, the increasing demand for integrated circuits, have led designers to outsource parts of the production process and use services provided by third-party vendors, which in turn exposes the production process to security concerns such as hardware Trojans. In this paper, we propose a machine-learning model utilizing Graph Convolutional Networks to detect hardware Trojans in gate-level netlists. In this regard, the proposed model, FAST-GO, is trained based on the extracted graph from the netlist and presents an efficient and compact set of structural gate-level features to classify the graph nodes into Trojan or normal nodes. Utilizing a small number of features in the presence of an efficient dataset enables the model to boost scalability and makes it a sensitive model applicable for large netlists. Experimental evaluation shows that our model has been successful in detecting 95.38% of Trojan nodes in 15 various-size circuits from the Trust-Hub benchmark only in 0.84 seconds on average using an Intel Core i5-3230M 2.6 GHz Processor with 6 GB RAM. This shows that FAST-Go is an accurate detection method with remarkably improved performance considering all known HT detection mechanisms.