Abstract Hiding secret data in digital multimedia has been essential to protect the data. Nevertheless, attackers with a steganalysis technique may break them. Existing steganalysis methods have good results with conventional Machine Learning (ML) techniques; however, the introduction of Convolutional Neural Network (CNN), a deep learning paradigm, achieved better performance over the previously proposed ML-based techniques. Though the existing CNN-based approaches yield good results, they present performance issues in classification accuracy and stability in the network training phase. This research proposes a new method with a CNN architecture to improve the hidden data detection accuracy and the training phase stability in spatial domain images. The proposed method comprises three phases: pre-processing, feature extraction, and classification. Firstly, in the pre-processing phase, we use spatial rich model filters to enhance the noise within images altered by data hiding; secondly, in the feature extraction phase, we use two-dimensional depthwise separable convolutions to improve the signal-to-noise and regular convolutions to model local features; and finally, in the classification, we use multi-scale average pooling for local features aggregation and representability enhancement regardless of the input size variation, followed by three fully connected layers to form the final feature maps that we transform into class probabilities using the softmax function. The results identify an improvement in the accuracy of the considered recent scheme ranging between 4.6 and 10.2% with reduced training time up to 30.81%.