2026·Prépublication
Serge Ebele, Stéphane Gaël R. Ekodeck, Chantal Marguerite Mveh-Abia, Batinyuy, Ngeh, Anna, Arthur Ulrich Ewane, René Ndoundam
HAL (Le Centre pour la Communication Scientifique Directe)
<div> Operational side channels remain a practical source of leakage in some high-capacity image steganography systems. WYSAWIS, a recent coverless indexing method, avoids pixel modification but relies on cloud-carried location information for extraction. We investigate a different operating point: CRC-verified payload recovery from one transmitted image without folder positions, block-index lists, or external location metadata. We propose ETHEGAN, an auxiliarychannel-free neural image steganography framework with a self-contained packet layer containing payload length, CRC32 validation, Reed-Solomon parity, deterministic interleaving, and optional reliability-aware erasure decoding. We further add residual-statistics regularization and a lightweight adversarial residual-domain discriminator to reduce visible and statistical embedding artifacts. On fixed 100-image Caltech101 and COCO val2017 clean-channel subsets at 0.25 nominal bpp, corresponding to 0.185546875 effective user bpp after packet and error-correction overhead, ETHEGAN recovers 100/100 Caltech101 payloads and 93/100 COCO payloads exactly after ECC and CRC validation. Increasing Reed-Solomon parity from 64 to 128 bytes per chunk lowers the effective user payload to 0.123046875 bpp and recovers 100/100 payloads on both datasets. Residual-strength and steganalysis-aware ablations preserve 50/50 exact recovery on both 50-image subsets while improving PSNR relative to the initial residual baseline. A lightweight residual/LSB detector remains effective over three grouped train/test splits, with mean test AUC 0.9835 on Caltech101 and 0.8763 on COCO. JPEG compression, blur, and resizing yield 0/20 exact recoveries in both datasets. A conservative low-payload detectability-gated mode at 0.001953125 effective bpp accepts 10/40 candidate images, recovers all accepted payloads, and yields test AUC 0.58 for the lightweight detector and 0.52 for a SPAM-style residual cooccurrence detector. An AES-256-GCM pre-packet encryption wrapper also passes correct-key, wrong-key, and wrong-associated-data authentication tests. The supported contribution is therefore clean-channel, auxiliary-channel-free payload recovery with effective-capacity accounting, optional confidential payload transport, and detector-bounded passive security only in the lowpayload gated mode; robustness remains unsupported. </div>
