DL-Steg: a deep learning-based steganography model for improving image security

Abstract The increasing social media and internet usage and augmentation of users are passing through an era of security risk as an increased concern. Steganography is the method of hiding secret messages inside digital mediums for the secure transmission of the messages. Steganalysis refers to the approach of detecting the presence of hidden information and retrieving it. An appropriate deep learning (DL)-assisted image steganography system should allow hiding and retrieval of hidden images with high accuracy and robustness, ensuring minimal changes in the quality of the cover image and security against detection. In this research work, the proposed framework is DL-based steganography framework integrating a Stacked Autoencoder (SAE) with Long Short-Term Memory (LSTM) networks for compressing, encoding, and embedding secret images into cover images in a secured manner. The evaluation is done on the Classified ImageNet dataset (a pre-processed version of the ImageNet dataset available on Kaggle where images are organized by class and resized for consistency in size). In the proposed method, DL-Steg initially preprocessed the cover and secret images. The encrypting of the secret image, after preprocessing, uses the Elliptic Curve Cryptography (ECC) technique, thus considerably maximizing its security. The secret image encrypted is given as the input to the SAE, which compresses and encodes the data efficiently. The LSTM was implemented to capture sequential dependencies in the image. The encoded secret image was fused with the cover images to produce a container image that can be securely stored or transmitted. These processes are reversed for the secret image’s decryption. DL-Steg has a maximum peak to signal noise ratio (PSNR) of 45.5940 and a structural similarity index measure (SSIM) of 0.9877, indicating better image quality and similarity and has a minimum mean squared error (MSE) of 0.143 and a good payload capacity of 24 bpp. The overall research model excelled against all the compared models with better results.