I am working on a project to develop a neural network for detecting steganography [messages hidden in image pixel values]. I am using a dataset available here, which contains 70,000 records of features extracteg from images histograms such as:

Kurtosis, Skewness, Std, Range, Median, Geometric_Mean, Hjorth Mobility, Hjorth Complexity, with tags (0/1)

I trained my network on this dataset and achieved 93% accuracy on the test set. Now, I want to create an application that takes an image as input, computes its histogram, extracts features, and then uses the trained model for classification.

Problem:
The features I generate from the histograms of new images differ significantly from the feature values in the training dataset. As a result, the neural network fails to classify new images correctly. I am not sure if the issue lies in the feature extraction process, preprocessing, or the neural network architecture itself.

Can anyone suggest how to correctly implement the feature extraction process or adjust the model to improve classification of new images?

-architecture is 4x Dense 256, 128,64 and 32 neurons aside with input=8 layer and sigmoid output neuron. I experimented with several layers and neurons. No success here

-Tried to add some dropout layers, nothing changed

-I tried to find some source cover images based on information given by dataset authors (‘Stanford Genome Project’ I don’t know, didn’t find any datasets related) to see why features differ and spot any mistakes in extraction.

-similar thing happened trying to make convolutional network trained on actual images

-alredy made some experimeting and confusion matrix, measures and stuff but it seems pointless now when it doesn’t have practical purpose ouside of training datasets.

For context, the goal is to make naural network and compare it to existing staganalysis tools for thesis. I would greatly appreciate any suggestions not only on how to solve the immediate issue but also on additional research, tests, or approaches that could enrich my work. I am using Python with Keras.