DK-CDS: Prompt-Sensitive Coverless Diffusion Steganography via Key-Guided Latent Interpolation

Diffusion-model-based coverless image steganography provides a promising way to hide visual information through generative models without modifying an existing cover image. However, existing methods still face two limitations: the semantics of the secret image may affect the generated stego image, and approximate prompts or partially matched access information may still reveal recognizable secret content. To address these issues, we propose DK-CDS, a prompt-sensitive coverless diffusion steganography method based on chaotic control and key-guided latent interpolation. In DK-CDS, the semantic access factor guides the visible content and style of the stego image, while the master key controls the chaotic sequence for latent-space access regulation. The secret image is implicitly embedded by establishing a matrix-weighted interpolation relationship between the secret latent variable and the prompt-guided natural latent variable. The prompt representation and the master key are further bound to the chaotic control process, making secret recovery highly sensitive to prompt variations and key mismatch. Under matched access factors, the secret image can be reconstructed through deterministic inversion and inverse matrix-weighted interpolation. Experimental results empirically demonstrate, under the evaluated settings, that DK-CDS generates visually natural and prompt-consistent stego images, achieves high-quality correct-access recovery, and effectively suppresses recognizable secret recovery under approximate prompts or mismatched keys. Among the evaluated representative coverless and diffusion-based steganography methods, DK-CDS obtains a higher task-specific auxiliary WCSI of 0.8356 and shows favorable performance in image quality, recovery fidelity, access sensitivity, steganalysis resistance, and robustness evaluations.