StegaCBAM-Net: An Attention-Driven Deep Steganographic Encoder-Decoder

The method of image steganography allows one to communicate secretly via encoding confidential information into harmless-looking images without compromising the visual appearance of the latter. The current approaches to steganography like LSB embedding and transform domain embedding provide low computational complexity, yet they lack robustness, have limitations in terms of payload, and are prone to steganalysis [1], [2]. Steganography using deep learning networks like CNNs has led to improvements in terms of both capacity and concealment quality as they can learn the embedding techniques through training on datasets [3], [4]. Yet most of the existing techniques use shallow features extraction and pay insufficient attention to spatial/channel dependencies. The proposed paper introduces an Attention-Based Image Steganography Model (AEDISA) to increase imperceptibility, payload, and robustness in image steganography systems. This framework uses an end-to-end learning scheme consisting of an encoder-decoder model using three neural networks, including a preparation network, hiding network, and revelation network. In addition, the model uses multi-scale convolution (3×3, 4×4, and 5×5) and a novel spatial-channel attention mechanism based on the CBAM approach for efficient learning of features. The model is trained end-to-end using the ImageNet100 benchmark dataset. Our experiment shows promising results compared to the state-of-the-art CNN steganographic approaches by obtaining 79.74 dB PSNR, 0.9703 SSIM, and higher reconstruction accuracy.