Gate-level Hardware Trojan Detection Method Based On Machine Learning

Hardware Trojan is one of the important security risks in the design and manufacture of integrated circuits.This is because the heavy reliance on third-party hardware IP(Intellectual Property)and automation tools in the design process,as well as the outsourcing of the design and manufacturing steps to external parties due to economic reasons,have led to the possibility of design in the manufactured integrated circuit chips.Malicious functions outside the specification lead to serious consequences such as data leakage or tampering.Hardware Trojan detection includes pre-silicon detection and post-silicon detection.The pre-silicon inspection is mainly for testing the design code,and the post-silicon inspection is mainly for testing the chips after manufacturing.They are at different stages and target different types of hardware Trojans.The former is a hardware Trojan that an attacker implants in the design code during the design process,and the latter is a hardware Trojan that is implanted in the chip during the chip manufacturing process.This thesis focuses on hardware Trojan that may be implanted in the code design process or when calling IP cores,and conducts a research on door-level Trojan detection methods based on machine learning.For different circuits,circuits containing Trojan are automatically generated for the circuit to be tested,and the feature values of the Trojan are extracted as the training set of machine learning,and finally,whether such hardware Trojan exist in the circuit to be tested is finally detected.1)This thesis proposes a method for automatically inserting hardware Trojan for any circuit.Based on the characteristics of hardware Trojan that are difficult to trigger,the proposed method implements hardware Trojan triggers and load circuits by inserting low-signal probability generating circuits and alternative selectors..At this time and at the same time,in order to ensure that the generated hardware Trojan circuit is difficult to be detected,the generated hardware Trojan circuit is simulated and filtered to obtain the available hardware Trojan circuit.Through the above method,multiple hardware Trojan circuits can be automatically generated for any circuit under test.2)This thesis is oriented to machine learning,and proposes two features of controllable rate and observable rate.For the training set,a variety of different input probabilities are adopted to obtain multiple sets of connection features,and then support vector machines are used to train them.The experimental results of ISCAS' 89 reference circuit show that the detection accuracy of hardware Trojan signal lines can reach 100%,while the detection accuracy of non-hardware Trojan signal lines can reach more than 99.93%.