Enhanced CNN Approaches for Multi-Image Embedding in Image Steganography

We unveil a convolutional neural network (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{CN N}$</tex>)-architectured steganographic model aimed at concealing multiple secret images in a single cover image, seeking to enhance payload capacity and minimize any encoding or decoding errors. We use a method that employs <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{CN N}{\mathrm{s}}$</tex> for encoding and decoding, harnessing a multi-scale encoder and a key-based decryption approach for increased security. Steganography is accurately and successfully achieved at two and three image levels by the model. For the use of two-image steganography, an accuracy of 95.16% was achieved for the encoded cover image, and decoding accuracies of 97.16% and 97.28% for secret images 1 and 2, respectively. In the three-image steganography, the accuracy of the coded cover image stood at 93.84%, while the decoded secret images achieved accuracies of 96.39%, 94.82%, and 94.76%. We boosted security by employing cryptographic techniques such as AES and ChaCha20 and instituted key integration at the architectural level. Our findings show the competent encoding and decoding of a range of secret images with enhanced security and noteworthy precision.