A method for detecting melamine in milk powder was developed based on near infrared (NIR) auto-correlation spectroscopy. Forty pure milk powder samples and forty adulterated milk powder samples with different relative factions of melamine (10-4%~40%, w/w) were prepared. The NIR reflectance spectra of all samples were collected in the range of 10 000~4 000 cm-1. Synchronous two-dimensional (2D) NIR correlation spectrum was calculated under the perturbation of melamine concentration, and the 7 000~4 200 cm-1 region was selected to establish a model. Then, based on the extracting information of auto-correlation spectra, the classification and quantification models of adulterated milk powder were established using partial least square (PLS) method. The 100% classification accuracy and the root mean square errors of prediction (RMSEP) of 0.63% were achieved, while the classification accuracy and RMSEP were 96.2% and 0.84%, respectively, using conventional one-dimensional NIR spectra, which showed that the auto-correlation spectra could provide better results, probably because more characteristic information could be extracted than conventional one-dimensional NIR spectra.