Deep Learning Evaluation of a Steganographic Algorithm

The embedding and transmission of secret information through multimedia such as images can be used to achieve privacy in medical and military applications. This form of covert communication is called steganography. As images are also used for training machine learning algorithms for medical diagnosis, it is important to evaluate if embedded data within images have adverse effects on the performance of medical machine learning models. This paper evaluates the effect of a previously developed steganographic algorithm called the
$C_{4}S$
algorithm on deep machine learning models for malaria diagnosis. The steganographic algorithm is used to hide patient data in malaria blood smear images. Both original and watermarked images are used to train different models using deep learning algorithms. The goal is to determine the effect of steganography on deep machine learning algorithms. The deep learning models are used to predict if a blood smear image has a malaria parasite or not. The results show that the basic Convolutional Neural Network (CNN) model trained with combined watermarked and non-watermarked images has the best prediction accuracy of 97.85%. Basic CNN trained with only non-watermarked medical images performed better on the watermarked test set with an accuracy of 97.65% as opposed to lower performance on a non-watermarked test set at 94.50%accuracy. We conclude that it is a false generalisation to assume that medical image steganography reduces the classification accuracy of machine learning algorithms. Instead, medical image steganography algorithms should be subjected to clinical trials and approved on a case-by-case basis for use in digital health applications