In order to overcome the shortcomings of the traditional support vector machines that the training samples should be labeled artificially this paper proposes an active training su pport vector machine model which uses affine propagation clustering algorithm to cluster unlabeled samples. The model does not need to manually label training samples,but updates the training data of existing suppo rt vector machines in the iterative process,so it can not only reduce the error caused by artificial mark ing samples,but also maximize the accuracy of the recognition model.In this paper,the terahertz spectra data of transgenic cotton are taken as the research object to verify the model.The recognition rate of the method is 95.56% for the total sample to be tested,which has less false and higher recognition rate than the other three methods. The results of experiment show that the support vector machine based on aff ine propagation clustering has higher recognition rate and lower misjudgment rate than the traditional support vector machine and it provides a fast and nondestructive method for the detection of genetically modified subst ances.