Mutual Normalized Garsia Wachs and Autoencoder Disentangle Log Likelihood Based Secure Steganography

Steganalysis is associated with the category of pattern recognition. Steganalysis mechanism has been evolved owing to the successful evolution of Deep Learning (DL). The objective of steganalysis remains in detecting whether cover image carries secret information that is being embedded by steganographic algorithms. The conventional steganalysis is trained on stego images created by steganographic algorithm, whose detection performance is said to be dropped swiftly when is being applied with another type of steganographic algorithm, therefore resulting discrepancy in steganalysis. To address on this issue, a deep learning driven feature-based method called Garsia Wachs Huffman Coding and Uniform Gradient Discriminative Auto-encoder (GWHC-UGDA) based secure steganography is proposed. Alternative to directly embedding the secret text message into a cover image, our GWHC-UGDA method hides it by transforming it into a synthesised image by means of Mutual Normalized Histogram and Garsia Wachs Optimal Huffman Coding and is hence inherently susceptive to typical steganalysis attacks. By disentangling an image into two representations for cover image and secret text message, Mutual Normalized Histogram Equalization and Garsia Wachs Optimal Huffman Coding-based Secret text message construction are applied separately to enhance steganography security. Following which an end-to-end steganylsis model employing Auto-encoder Disentangle Uniform Gradient Discriminative Log Likelihood-based Steganography is designed that performs both the encoding and decoding concurrently. Extensive experiments are performed to reveal the efficiency of the proposed GWHC-UGDA method. We explore that the reveal that the GWHC-UGDA is superior to the compared state-of-the-art deep-learning secure steganography methods in terms of PSNR, bit error rate, extraction error and extraction accuracy. In addition, the proposed GWHC-UGDA method has good attributes for image security that in addition improves the performance of the proposed scheme of steganography in terms of extraction error by 51% and improving extraction accuracy by 26% respectively.