Advanced glycation end products (AGEs) are a complex and heterogeneous group of compounds that have been implicated in diabetes related complifications. Skin autofluorescence was recently introduced as an alternative tool for skin AGEs accumulation assessment in diabetes. Successful optical diagnosis of diabetes requires a rapid and accurate classification algorithm. In order to improve the performance of noninvasive and optical diagnosis of type 2 diabetes, support vector machines (SVM) algorithm was implemented for the classification of skin autofluorescence from diabetics and control subjects. Cross-validation and grid-optimization methods were employed to calculate the optimal parameters that maximize classification accuracy. Classification model was set up according to the training set and then verified by the testing set. The results show that radical basis function is the best choice in the four common kernels in SVM. Moreover, a diagnostic accuracy of 82.61%, a sensitivity of 69.57%, and a specificity of 95.65% for discriminating diabetics from control subjects were achieved using a mixed kernel function, which is based on liner kernel function and radical basis function. In comparison with fasting plasma glucose and HbA1c test, the classification method of skin autofluorescence spectrum based on SVM shows great potential in screening of diabetes.