Hiding data in images using deep neural networks.

Image based Steganography has been in existence for a very long time and Kurak and McHugh were one of the earliest who proposed a method in which the secret data or image is embedded into the fourth least significant bit. But this method was not of much use since the intruder can easily access the hidden image or data since they know exactly the position of the hidden data. This is an example to show that this is an area in which security and efficiency can be improved by using different methodologies. This paper is proposed to build an architecture to hide single or multiple images also called hidden images which are the inputs to be hidden inside an outer image and then the combined image is converted into an Stego image, and the image looks like the outer image except that it is filled with the bits of information from the hidden image. From this Stego-image can be retrieved the original outer image using disclosure component which is the reveal network. A deep learning model has been developed and trained which consists of two layers namely preprocessing layers and embedding layers which will hide the data or image inside the outer image and convert it into a Stego image. Then from this Stego image the hidden data is extracted using reveal network. Only this reveal network can extract the data and no intruder can extract it since the embedding is done by the model and it does not use same or fixed pattern. This method helps to transmit sensitive information in a more secure and efficient manner. The extracted image from the outer image has an accuracy of 95 percent only for the first image and it decreases for each added hidden image and the MSE value is 0.01 and it increases for increasing number of hidden images. The model has a F1 score of 0.96. [ABSTRACT FROM AUTHOR]