Title: Training-Free Coverless Multi-Image Steganography with Access Control

Abstract:

Coverless Image Steganography (CIS) offers enhanced imperceptibility and inherent resistance to steganalysis by embedding data without directly altering the cover image. Despite these advantages, current CIS approaches generally fail to provide robust access control, complicating the task of selectively disclosing distinct hidden messages to specific authorized recipients. Implementing such control is essential for managing privacy and scalability in multi-user information hiding scenarios. To address this, we introduce MIDAS (Multi-Image Diffusion-based Access-controlled Steganography), a novel framework that leverages diffusion models to enable coverless steganography without requiring training. MIDAS facilitates the concealment of multiple images and allows for user-specific access control through latent-level fusion. The system incorporates a Random Basis mechanism designed to minimize residual structural artifacts, supported by a theoretical examination of information leakage. Additionally, it features a Latent Vector Fusion module that transforms combined latents to better synchronize with the diffusion process. Our experiments indicate that MIDAS surpasses existing training-free CIS baselines across several key metrics, including access control capabilities, the quality and diversity of stego images, resilience against noise, and evasion of steganalysis. These findings position MIDAS as a practical and scalable solution for access-controlled coverless steganography.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC