Quality and authenticity of milk are usually defined by the content and composition of fat, proteins, total solids, somatic cells and other established indicators. Along with traditional reference methods for quantitative analysis, the use of near infrared (NIR) spectrometers and multivariate modelling is becoming more widely utilised. It would be useful to apply a spectrometer with a linear silicon CCD array as a rapid, portable and rather cheap supervising device. However, milk is a multiphase disperse system which includes a rough disperse phase of somatic cells and fatty globules, a thin phase of casein and whey proteins particles and the molecular solutions of lactose, salts, vitamins etc. in water. The scattering of incident light by particles of milk affect the signal of an NIR spectrometer significantly. The calibration of a spectrometer can be improved by considering the distinction between scattering by fatty globules and casein micelles. For this purpose, it is attractive to use the anisotropy of scattering by rough particles, in particular, by fatty globules. The comparative results for calibration models constructed with transmission and back scattering spectra of milk are presented in this paper. The spectra were acquired with a new two-channel short wave NIR spectrometer. Restored drinking milk has been modelled with a set of calibration samples, since the difference in the optical properties due to the globules and the micelles is significant and also because the reliable quality control of drinking milk using a simple spectrometer is a relevant practical problem. PLS models were constructed for the prediction of fat and protein and also of the fat-free total solids which serve as an indicator of authenticity for drinking milk. The results obtained demonstrate that a large reductions in errors is possible. The SEP value of not more than 0.08% wt was achieved when predicting the fat content of milk using a calibration based on the difference spectra of transmission and backscattering compared with a SEP of 0.21% wt for the model using transmission spectra only. For protein determination, the SEP was reduced from 0.25% wt to 0.12% wt if the calibration was based on the spectra of backscattering rather than on transmission.