International Journal of Science, Engineering and TechnologyAuthors: Mr. Somnath Pandurang Gunjal, Dr. Sanjay Kumar Sen
Abstract: The rapid growth of digital communication and image-based information exchange has increased the need for secure methods for protecting confidential visual content. This paper presents DeepSafe, a deep-learning-based image steganography framework designed to conceal a secret image inside a cover image while maintaining visual imperceptibility and enabling subsequent recovery. The framework is based on three convolutional neural network (CNN) components: a Preparation Network that transforms the secret image into learned feature representations, a Hiding Network that combines the prepared secret representation with the cover image, and a Reveal Network that reconstructs the hidden image from the generated container image. The original implementation uses RGB images resized to 224×224 pixels, normalization by 255, the Adam optimizer, mean squared error (MSE) loss, 100 training epochs, and a batch size of 12. The reported dataset comprises images obtained from Kaggle and cover images from the ALASKA steganography dataset, with 500 data samples described in the original implementation. The revised paper clarifies the workflow, network configuration, experimental protocol, security considerations, attack scenarios, and limitations. The source manuscript reports evaluation using accuracy, precision, recall, F1-score, and AUC and states that the proposed approach improves image-security performance; however, the underlying numerical values of the plotted results are not available in the supplied manuscript. Accordingly, this revision does not fabricate numerical results. DeepSafe is positioned as a foundation for secure communication, digital forensics, privacy-sensitive data exchange, and future adversarially robust image-protection systems.
DOI: http://doi.org/10.5281/zenodo.23188205
