Detection Methods of Hardware Trojans Based on DBN and QPSO-OCSVM Anomalous Behavior Analysis

Abstract At present, hardware trojan detection methods based on the side channel analysis technology mainly depend on the presence of the template chip, resulting in a problem that it is difficult to put the methods into practical applications. In addition, detection focus is always paid on the impact of the hardware trojan on the target circuit, and impact of hardware trojan trigger state on the target circuit is ignored. To address the abovementioned problems, a hardware trojan detection method based on anomalous behavior analysis is proposed in this paper. Firstly, detection theory based on anomalous behavior analysis is proposed, and an anomalous behavior analysis methods based on the DBN (Deep Belief Networks) and the QPSO-OCSVM (Quantum Particle Swarm Optimization-One Class Support Vector Machine) is designed, based on which a detection model is established. This paper first proposed the detection theory based on anomalous behavior analysis, designs the anomalous behavior analysis methods based on DBN and QPSO-OCSVM, and build a detection model on this basis. Then, a hardware trojan detection platform designed and implemented, where three trigger-type hardware Trojans for the AES(Advanced Encryption Standard) encryption in the international Trust-Hub library are selected as the detection object and implanted to the encryption chip. Finally, the DBN model is utilized to extract the feature and reduce the dimensionality of the captured behavior data. The QPSO-OCSVM is used for detecting the anomalous behavior and obtaining the final detection result. Experimental results show that the detection method proposed in this paper could achieve the detection accuracy higher than 99% for the hardware trojans with resource occupancy rates of only 0.33%, 0.34%, and 0.76%. Optimal analysis model is established and a performance comparison with the anomalous detection model based on Isolation Forest and Local Outlier Factor is implemented, exhibiting that the detection model proposed in this paper has the highest performance.