Abstract: Hardware Trojan detection is too difficult due to the reason that chip production chain is too long. An improved machine learning classification algorithm is proposed. First, the time delay signals of circuits under different voltages are collected, and then determine whether there is a Trojan horse in the circuit depending on the time delay. A polynomial regression algorithm is combined to fit the delay data, and Trojan horse feature library can be established based on the regression function. The experimental results show that the 19 different voltages of 2 000 groups of delay units are extracted to compare. When the number of voltages considered comprehensively is 19, its prediction accuracy reaches the highest 95.2%. The proposed classification and regression algorithms can improve the recognition accuracy and automation of the hardware Trojan.