To detect adulterated milk rapidly and accurately, the discrimination models for adulterated milk were established based on the method of Kernel Orthogonal Projection to Latent Structure (K-OPLS). By using the Gaussian radial basis function as the kernel function and the minimum value of Root Mean Square Errors of Cross-validation (RMSECV)as an evaluation index, the width of the Gaussian kernel, the minimum value of the RMSECV, and the number of Y-orthogonal components (scalar) were selected in a optimization. 36 samples with different concentrations of tetracycline (0.01-0.3 gL-1), melamine (0.01-0.3 gL-1) and glucose (0.01-0.3 gL-1) in milk were prepared, respectively. Then the infrared absorption spectra of all samples were measured. K-OPLS models for tetracycline-tainted milk, melamine-tainted milk and glucose-tainted milk were constructed. The results show that its classification accuracy for tetracycline-tainted milk, melamine-tainted milk and glucose-tainted milk are 100%,100%,95.8%, respectively. The K-OPLS model was used to classify the above three kinds of adulterated milk and pure milk and its classification accuracy for unknown samples is 93.1%. Compared with Partial Least Square Discriminant Analysis (PLS-DA), K-OPLS methods show higher accuracy. The results indicate that the K-OPLS model has good prediction ability for complex milk systems.