Blind Content-Adaptive Multi-Image Steganography Using Canonical Sparse Multi-Transform Residual Maps and AES-GCM
Abstract
Context: Multi-image LSB steganography can reduce payload per cover image, but conventional content-agnostic allocation and non-blind extraction limit practical use and do not adequately address modern steganalysis. Objective: This work develops a blind, lossless-delivery multi-image steganography method that uses canonical sparse multi-transform residual maps to determine payload allocation and embedding locations. Method: Each cover is converted to an LSB-normalized canonical image, (C_i=I_i\ \mathrm{AND}\ 254), which remains invariant after one-LSB substitution and enables the receiver to reconstruct the DWT/DCT residual maps, residual-energy allocation, and keyed embedding schedule from the received stego images alone. AES-256-GCM provides authenticated encryption, while a collision-free keyed bootstrap header conveys the required protocol information and a fresh 96-bit nonce. Experiments used a publicly available 9,000-image, (256\times256) grayscale BOSS-derived collection, arranged as 4,500 two-image transmissions and split at group level into 80% training, 10% validation, and 10% test groups. Results: At requested rates of 0.05, 0.10, 0.20, and 0.30 bpp, all 36,000 group-method-rate executions achieved canonical equality, successful AES-GCM authentication, exact plaintext recovery, and zero observed sealed-payload BER. At 0.05 bpp, CSR-LSB achieved (64.49\pm1.87) dB PSNR and (0.999888\pm0.000105) SSIM on 900 held-out test images. It reduced compact residual-feature detector AUC from 0.9978 to 0.8514 and deep residual CNN AUC from 0.9957 to 0.9366 relative to uniform keyed LSB. Conclusions: CSR-LSB provides blind authenticated recovery and lower detectability than uniform keyed LSB, although modern detectors still identify its stego images reliably; future work should incorporate distortion-minimizing embedding to improve resistance to advanced steganalysis.
