In recent years, deep learning has made a series of breakthroughs in processing near-infrared spectroscopy, Raman spectroscopy, fluorescence spectroscopy and other spectroscopy data. However, due to the high demand for deep learning methods for the size of the training set, and it is difficult to obtain a large number of labeled samples in the field of analytical chemistry, the overfitting issue has always been highly-concerned by researchers in the application of deep neural network in chemometrics. In response to this problem, this paper proposed a near-infrared spectra modeling method based on the wavelength attention-convolutional neural networks (WA-CNN) and applied it to the quantitative analysis of hydrolyzed leather protein (HPL) adulteration in infant formula. WA-CNN adds a wavelength attention module based on the traditional convolutional network. This module uses convolution operation to learn the attention weights and activates the effective bands in the form of multiplication, thereby effectively alleviating the redundancy and over-fitting problem in NIR modeling based on deep learning. A total of 100 HLP adulterated infant formula samples were tested, and the adulteration ratio was in the range of 0% to 20%. Random sampling was performed 10 times for modeling, in which 60% of the samples were used for training, while the remaining 40% of samples were adopted for testing. The model was evaluated by the mean of root mean square error (RMSEP), coefficient of determination (R2) and relative analysis error (RPD). Three traditional models, namely partial least squares regression (PLS), support vector machine regression (SVR) and conventional one-dimensional convolutional neural network (CNN), were also established for comparison. Compared with the above comparison methods, WA-CNN achieved the best model prediction results and obtained RMSEP=1.32%±0.12%, R2=0.96±0.01, RPD=4.92±0.41. In addition, the results also show that the WA-CNN has a faster and more stable convergence process than the traditional CNN for both the training set and the test set during the training process. Moreover, in the case of different training sample sizes (ranging from 20% to 80%), WA-CNN also achieves the best accuracies among all the examined models.