Work

Diffusion-enhanced DCT steganography: reward-guided cover generation for increased embedding capacity with exact lossless-channel recovery

Year: 2026
Type: preprint
Abstract: Discrete cosine transform (DCT)-domain steganography embeds data into quantised DCT coefficients, but its capacity depends on how many cover coefficients are large enough to carry a bit. We propose Diffusion-Enhanced DCT Steganography (DEDS), which fine-tunes a generator to produce covers with more embeddable coefficients. Stable Diffusion v1.5 is fine-tuned with Low-Rank Adaptation, a differentiable DCT reward and an anchor loss. The payload is written by a margin-projected parity embedder whose expected distortion, ( 1/2 +μ)3Q2 / 3, and guaranteed coefficient-domain tolerance, μQ, we derive. On 1,000 prompt- and seed-matched 256 × 256 cover pairs, fine-tuning raised the mean number of eligible coefficients per 8 × 8 block from 1.57 to 16.64 and raw slot capacity from 0.056 to 0.418 bits per pixel (bpp) on the 566 pairs with cover-derived budgets. At 0.30 to 0.60 bpp every message was recovered exactly, with a luminance peak signal-to-noise ratio (PSNR) of 39.9 dB on fine-tuned covers, 6.8 dB above bin-centre embedding and above baseline covers in every payload bin. An optional redundancy mode survived mild JPEG compression and noise at up to 0.02 bpp. Rich-model steganalysis caught the four content-adaptive schemes (area under the curve, AUC, 0.85 to 0.99) and DEDS (AUC 0.997), so the method’s value lies in capacity and fidelity, not security.
Institution: BRAC University
Language: English
Open Access status: green
OpenAlex ID: W7234348549
Cites: 42
Cited by:
Related to: 10