This paper describes a method for determining whether a digital image transmitted over a computer network has been encrypted to hide a message in cover carriers (pictures), so it cannot be detected on a computer network accessible to those for whom the message is not intended.
Normally the "noise" of an original digital image is the Gaussian noise. Discovery that the image noise is not normally distributed can be the first evidence that the original image has been encrypted. In order to detect the distribution of the image noise, the special neural network method has been developed and tested. The method of the Gaussian distribution detection by the neural networks uses approximation techniques to detect the shape of the noise distribution. The new method of noise distribution shape detection uses two types of approximation of the sampled noise amplitude distribution. The first method is called "model-les" and uses the feed-forward neural network to perform the approximation. The second method is the classical least square approximation that uses the given Gaussian function as the model of approximation. The best fit of the Gaussian function is found to be the sampled noise amplitude distribution. The comparison of the two approximations shows whether the noise in the picture is normally distributed. In the normal noise distribution, both approximations produce nearly the same results. 1 table, 3 figures, mathematical equations, and 4 references