Abstract

With the tremendous growth in the use of information technology, the connected health networks are becoming more relevant, greatly improving the traditional standard of healthcare procedures from data acquisition, storage, and sharing among the medics for timely clinical diagnosis processes, therapy, and disease management. However, connected health network comes with network and cyber criminality challenges, and frequent security breach attacks on digital platforms and databases. Unfortunately, Sensitive clinical information is greatly at risk with adversaries, most clinical stakeholders find it difficult to overlook free access to the clinical records. Previously, feature aggregation networks, convolutional neural networks, residual convolutional neural networks, and machine learning models were used in different methodological approaches towards ensuring the detection of hidden information in images, unfortunately, none was able to produce optimal results. This study proposes a three-phased framework to determine the suitability of embedder networks’ feature extraction for image steganalysis, predicting and detecting hidden information in images. A Multilayer Perceptron (MLP) deep learning model was trained for pattern recognition of steganography instances in acquired digital image signals. The digital image signals used for the predictive steganalysis are publicly available images contained in two circumstances highlighted regarding clean image signals (situation of cover images, but without steganography) and the embedded situation of image signals (where images (stego) with hidden data or information). Interestingly, the results of the parameter show that the Max-iter parameter of the MLP classifier hugely determines the performance of the algorithm towards detecting steganography in digital image signals. The parameter stipulates the number of times the training set will pass through the MLP network for the training process. Significantly, in our experiment, Max-iter returned the best result at 1000 netting an accuracy of 93%, precision of 57%, and recall of 100% weighted averages. Our study does not only implement a model that detects hidden information in images, but it also discovered and tuned the multilayer perceptron to determine where it will perform best.