With the further development of terahertz technology, terahertz has shown its unique advantages in food safety detection. Flour (wheat flour) is the staple food in most areas of northern China. Besides, benzoic acid(BA), as the important preservative of acid food, is often added to extend the preservation time of food. However, the excessive use of food additives would cause serious damage to human health. This paper explores the feasibility of detecting food additives through terahertz technology and conducts empirical study on benzoic acid in flour by terahertz time-domain spectroscopy (THz-TDS) technology. The terahertz time-domain and frequency domain spectrum of the mixed samples (flour and benzoic acid) were obtained. As shown by absorption coefficients, benzoic acid presented obvious absorption peak at 1.94 THz. Meanwhile, the absorption coefficient of flour increased at a certain slope, which indicated that the characteristic identification of benzoic acid in flour could be carried out by terahertz technology. In order to establish the quantitative detection model of benzoic acid additive in flour, terahertz time-domain spectra of benzoic acid doped with different percentages (mass fraction) in flour were collected, and the absorption coefficient spectrum was obtained through calculation. It was found that the absorption peak amplitude enjoys positive correlation with benzoic acid content. As for the detection method, firstly, explore the effects of different spectral pretreatment methods on THz spectroscopy, and then adopt methods like Smoothing, Multiple Scatter Correction (MSC), Baseline and Normalization to carry out correct processing. After correction, PLS model was established to select the optimal pretreatment method. Secondly, establish PLS and LS-SVM regression models for the determination of benzoic acid content in flour. The experimental results verify that PLS model established after normalization was more optimal, with correlation coefficient of prediction (rp) of 0.979 and root mean square error of prediction (RMSEP) of 1.30%. By comparison, it was proved that the most optimal quantitative determination model of benzoic acid content in flour is LS-SVM model with correlation coefficient of prediction (rp) of 0.987 and root mean square error of prediction (RMSEP) of 1.10% after the normalization of terahertz absorption coefficient. MLR model was established by only two bands of 1.946 and 1.869 THz with correlation coefficient of prediction (rp) of 0.955 and root mean square error of prediction (RMSEP) of 1.90%. It is concluded that a new solution for the nondestructive detection of benzoic acid additives in flour was developed, and method guidance was provided for the detection of other types of additives, all of which have an important significance for the healthy development of flour industry.