Optimized steganographic embedding guided by snake algorithm and fusion-aware attention maps

Image steganography aims to conceal secret data within digital images while preventing detection by both human observers and steganalysis. An effective steganographic system must satisfy three key requirements: imperceptibility, sufficient embedding capacity of Bits Per Pixel (BPP), and robustness against common image processing operations, such as compression and noise. This paper proposes a novel attention-guided steganography framework that employs Spatial Resolution (SR) and Histogram Equalization (HE), and uses an improved Snake Optimization Algorithm (SOA) to adjust the image distortion positions. The proposed steganography system uses an adaptive Least Significant Bit (LSB) substitution method. Unlike existing steganography methods that embed the same number of bits at the same complexity to the entire image, different bits (from 1 to 4 bits/pixel) can be adaptively embedded into the attention region or the non-attention region based on different attention values. In addition to implementing compulsive mechanisms for avoiding visual and spatiotemporal checks, our method also offers the option of embedding a mechanism based in Advanced Encryption Standard Galois Counter Mode (AES-GCM) for payload encryption and protection. Experimental results demonstrate that proposed steganographic method is visually imperceptible, achieving high Peak Signal-to-Noise Ratio (PSNR) ranging from 54 decibels (dB) to 61 dB, as well as a high Structural Similarity Index Measure (SSIM) value over 0.99998. Additionally, the method enables complete lossless recovery of hidden secrets and is hard to detect by steganalysis, achieving Area Under the Curve (AUC) in the range of 0.49-0.53, using existing CNN-based steganalysis networks (i.e., Xu-Net and Ye-Net).