Answers are grounded in this paper and cite their sources.
Image steganography must preserve image fidelity while limiting exposure of both hidden content and the hiding model. We present CIIS-Net, a key-conditioned camouflage-integrated invertible network that performs public image denoising without a key and secret-image hiding and recovery with the correct key. Its hidden-weight parameterization combines deterministic key-derived values with trainable residuals. Public pretraining is followed by joint optimization of public and secret parameters under a routine-preservation objective. Hiding and recovery share the forward and inverse directions of a Haar-wavelet affine coupling network. Across 250 cover–secret pairs from DIV2K, COCO, and ImageNet, the model achieves mean steganogram (stego) and secret-recovery peak signal-to-noise ratios (PSNRs) of 43.25 and 36.06 dB, respectively, under an eight-bit channel. Public denoising reaches 32.24 and 31.58 dB on 500 images at noise standard deviations of 15/255 and 20/255. Across three training runs, 1000 incorrect keys per run yield mean recovery PSNRs of 8.38–9.70 dB, demonstrating sensitivity to the evaluated key substitutions rather than formal security. On 1000 held-out cover/stego pairs, the evaluated StegExpose configuration yields an area under the receiver operating characteristic curve (AUC) of 0.5101, indicating near-chance discrimination. Yedroudj-Net yields an AUC of 0.9973 and 97.60% detection accuracy after training on 150 pairs for 150 epochs.
No takes yet. Share an insight, caveat, or question.
Eryılmaz et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: