Near-infrared spectroscopy was used to study the mixed solution of transgenic oil/non-transgenic oil. The acquired original spectra were pretreated with multiple scattering correction (MSC), first-order derivative (FD), moving window smoothing (MWS) and savitzky-golay smoothing first-order derivative (SG1). The effects of different pretreatment methods on the discriminant analysis of transgenic oil/non-transgenic oil support vector machine (SVM) modeling were compared. The model after MSC pretreated has the best prediction effect, and the accuracy rate is 91.6%. In order to further improve the accuracy and stability of the model, the successive projections algorithm (SPA) is used to select the characteristic band. We input the 15 feature wavelengths after SPA selection into the SVM, and the prediction accuracy is 98.3%. The experimental results show that the near-infrared transmission spectrum combined with stoichiometry can achieve rapid and non-destructive detection of transgenic oil/non-transgenic oil, not only suitable for the identification of pure genetically modified oils, but also for the identification of non-transgenic oils that are mixed with genetically modified oils.