Accurate Latent Inversion for Generative Image Steganography via Rectified Flow

2025-08-01

AI Summary

To address inaccurate latent variable inversion in diffusion-based steganography—leading to failed secret information extraction—this paper proposes RF-Stego, a high-fidelity steganographic framework leveraging path consistency and Rectified Flow (RF). Methodologically, we introduce PCLI (Path-Consistent Linear Inversion), a linear inversion process that rigorously aligns forward generation and backward reconstruction trajectories in latent space. We theoretically prove that Rectified Flow guarantees invertibility and numerical stability in the latent space, and design a lightweight RF sampler for efficient embedding and extraction. Experiments demonstrate that RF-Stego consistently outperforms state-of-the-art methods across all key metrics: extraction accuracy, image fidelity (FID reduced by up to 28%), robustness against noise, security against steganalysis, and generation efficiency (50% fewer sampling steps).

Technology Category

• Computer Vision: Diffusion Models for Vision
• Reasoning under Uncertainty: Stochastic Optimization
• Machine Learning: Privacy

Application Category

• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

Steganography based on diffusion models has attracted increasing attention due to its ability to generate high-quality images and exhibit strong robustness. In such approaches, the secret message is first embedded into the initial latent variable, and then the stego image is generated through the forward process. To extract the message, an inversion process is required to reconstruct the latent variables from the received image. However, inaccurate latent inversion leads to significant discrepancies between the reconstructed and original latent variables, rendering message extraction infeasible. To address this issue, we propose extbf{RF-Stego}, a novel generative image steganography method that enables accurate latent inversion and significantly improves extraction accuracy. First, we develop the extbf{P}ath extbf{C}onsistency extbf{L}inear extbf{I}nversion ( extbf{PCLI}), which imposes formal constraints on the inversion process. By explicitly aligning it with the forward generation path and modeling both directions along a shared linear path, PCLI eliminates path mismatch and ensures path consistency throughout the steganographic process. Second, through rigorous theoretical proof, we demonstrate that extbf{R}ectified extbf{F}low extbf{(RF)} offers both theoretical reversibility and numerical stability in the inversion process. Based on this, we replace traditional unstable samplers with RF sampler which effectively improves the numerical precision of the inversion process. Experimental results show RF-Stego outperforms state-of-the-art methods in terms of extraction accuracy, image quality, robustness, security and generation efficiency.

Problem

• Inaccurate latent inversion hinders message extraction in steganography
• Path mismatch in inversion reduces steganographic process consistency
• Traditional samplers lack stability for precise latent variable reconstruction

Innovation

• Path Consistency Linear Inversion for alignment
• Rectified Flow ensures reversibility and stability
• RF sampler improves numerical inversion precision