The work aims to establish an efficient and accurate research method for infrared spectroscopy inspection of paper packaging boxes. Infrared spectral data of 56 paper packaging boxes of different types and sources were measured with a Fourier transform infrared spectrometer. According to the different main fillers of the paper box, the samples were preliminarily divided into three categories. Principal component analysis was conducted to reduce the dimensionality of the preliminarily classified data. Four principal components were extracted, and the samples were finally divided into six groups through systematic clustering. The K-means clustering algorithm combined with Pearson’s chi-squared test was used for validation. The results were basically consistent with the final classification. Based on this grouping, three discriminative models, namely Naive Bayesian classification, random forest model, and XGBoost classification, were trained to achieve classification prediction for new sample groups. The 56 samples of express paper packaging boxes were divided into 3 categories and further subdivided into 6 groups. All three discriminant prediction models had high accuracy, with the random forest model having the highest accuracy. This method quickly and conveniently distinguishes samples, and can achieve non-destructive testing, providing a basis for the identification of paper packaging boxes at crime scenes, thus providing assistance for the development of public security investigation work.