High Payload Image Steganography Using DNN Classification and Adaptive Difference Expansion

The primary intention of this study is to enhance the capacity while ensuring visual integrity and maintaining privacy. A novel information-hiding method based on Adaptive embedding and Machine Learning classifier is proposed. To improve the capacity, the private message is compressed using Adaptive Huffman Encoding. The cover image blocks are classified based on the collected features from each block, determining the ideal Embedding Capacity (EC). The suggested approach has been verified in three distinct categories: the effectiveness of ML classifier, embedding method efficiency, and its robustness. DNN block classifier outperforms the LR and RF classifiers with 99% validation accuracy. The embedding approach with DNN as the block classifier got a PSNR close to 50 at an embedding rate of 1.22 bpp. The robustness of the proposed method is demonstrated with random addition of S&P to the Stego image at 0.5bpp of EC. The proposed scheme sustain S&P noise with accuracy 77% for 50% noise density.