Simultaneous species origin authentication and adulterant quantification in specialty milk powders using near-infrared spectroscopy and multi-task deep learning.

The adulteration of high-value specialty milk powder with low-cost bovine milk poses a serious challenge to food safety and authenticity. To address this issue, this study proposes a unified end-to-end multi-task learning multilayer perceptron (MTL-MLP) framework. This framework directly maps collinear spectral features to a deep shared latent space with minimal spectral signal preprocessing, and employing a novel Concentration-Conditioned Loss Optimization (CCLO) dynamic masking mechanism, the model successfully overcomes the "matrix swamping effect" under severe adulteration. As corroborated by t-distributed Stochastic Neighbor Embedding, the framework effectively isolates and identifies the pure bovine milk samples adulterated into the three specialty matrices. Quantitative results demonstrate that the MTL-MLP achieved an overall matrix classification accuracy of 93.10% across the entire concentration gradient. Notably, the model attained 100% authentication accuracy within the critical core identification zone (≤50% adulteration). The framework exhibited superior regression precision in quantitative adulteration tracking, yielding a coefficient of determination (R2) of 0.9987 and a root mean square error (RMSE) of 6.87 × 10-3. This NIR-based multi-task learning approach reliably detects cow milk adulteration down to a tested concentration of 1%, proving its feasibility for rapid on-site industrial screening.