Optimized Steganographic Embedding Guided by Snake Algorithm and Fusion-Aware Attention Maps

Image steganography involves concealing secret data within digital images so that no one can detect the presence of the hidden message. A steganography system requires three essential elements to succeed: (i) the embedding process needs to remain invisible to human observers and statistical detection systems, (ii) the system needs to maintain enough storage capacity (Bits Per Pixel (BPP)), and (iii) the system needs to withstand regular image alterations, such as compression and noise while ensuring successful message retrieval. Our research introduces an adaptive Least Significant Bits (LSB) steganography system that uses attention guidance to balance its performance goals. The Spectral Residual (SR) saliency method produces an attention map that can be combined with Histogram Equalization (HE) through multiplication to focus on areas with textured patterns that remain perceptible. Next, we apply the BPP policy, which receives its values from pixel attention quantization, producing values between 1 and 4 bits. The Snake Optimization Algorithm (SOA) then performs lightweight optimization to find the best pixel locations for embedding by optimizing the total attention value within given capacity constraints. Moreover, the system includes an optional Advanced Encryption Standard Galois Counter Mode (AES-GCM) encryption mechanism to protect both the payload data and its integrity. The system demonstrates excellent cover image preservation, with Peak Signal-to-Noise Ratio (PSNR) values ranging from 54 to 61 dB and Structural Similarity Index Measure (SSIM) values exceeding 0.99998, while achieving perfect secret extraction via lossless recovery. The proposed system achieves a near 50\% accuracy in Convolutional Neural Network (CNN) steganalyzers (Xu-Net, Ye-Net) across all payload levels, with AUC values ranging from 0.49 to 0.53. The proposed system also balances its three main performance factors through its optimized process, reproducible optimization, and optional cryptographic protection features.