Colorectal cancer (CRC) is the third cause of cancer-related death and the fourth most frequently diagnosed cancer across the globe. The objective of this study is to obtain novel and effective diagnostic markers to enrich CRC diagnosis methods. Herein, exosomal miRNA expression data of CRC and normal blood were subjected to XGBoost algorithm, and 5 miRNAs related to CRC diagnosis were primarily confirmed. Then multilayer perceptron (MLP) classifiers were constructed based on different subsets. Via integrated feature selection (IFS), we noticed that the MLP classifier constructed by the first four miRNAs (miR-654-5p, miR-126, miR-10b, and miR-144) had the highest Matthews correlation coefficient (MCC). Subsequently, principal component analysis (PCA) for dimensionality reduction was performed on samples based on the miR-654-5p, miR-126, miR-10b, and miR-144 expression data. The signature based on these four feature miRNAs, as the analysis indicated, could effectively distinguish CRC samples from normal samples. Further, we extracted the exosomes from clinical blood samples and applied qRT-PCR analysis, which revealed that the expression of these four feature miRNAs was in the trend of that in the test set. Collectively, these four feature miRNAs might be tumor biomarkers in the serum, and our study offers innovative thinking on early-stage CRC diagnosis.