SMILENet: Unleashing Extra-Large Capacity Image Steganography via a Synergistic Mosaic InvertibLE Hiding Network

TL;DR

SMILENet introduces a synergistic mosaic invertible hiding network to push image steganography beyond traditional capacity limits. By combining a non-reversible Secret Information Selection with reversible ICDM/IMSE modules, it forms a Mosaic Secret Representation that minimizes interference among many secret images while enabling accurate recovery. A new Capacity-Distortion Trade-off metric provides a principled evaluation of capacity versus distortion across varying secret-image counts. The approach achieves up to 25 hidden images with high fidelity and demonstrates strong anti-steganalysis performance, marking a significant advance in practical, high-capacity steganography with manageable complexity.

Abstract

Existing image steganography methods face fundamental limitations in hiding capacity (typically $1\sim7$ images) due to severe information interference and uncoordinated capacity-distortion trade-off. We propose SMILENet, a novel synergistic framework that achieves 25 image hiding through three key innovations: (i) A synergistic network architecture coordinates reversible and non-reversible operations to efficiently exploit information redundancy in both secret and cover images. The reversible Invertible Cover-Driven Mosaic (ICDM) module and Invertible Mosaic Secret Embedding (IMSE) module establish cover-guided mosaic transformations and representation embedding with mathematically guaranteed invertibility for distortion-free embedding. The non-reversible Secret Information Selection (SIS) module and Secret Detail Enhancement (SDE) module implement learnable feature modulation for critical information selection and enhancement. (ii) A unified training strategy that coordinates complementary modules to achieve 3.0x higher capacity than existing methods with superior visual quality. (iii) Last but not least, we introduce a new metric to model Capacity-Distortion Trade-off for evaluating the image steganography algorithms that jointly considers hiding capacity and distortion, and provides a unified evaluation approach for accessing results with different number of secret image. Extensive experiments on DIV2K, Paris StreetView and ImageNet1K show that SMILENet outperforms state-of-the-art methods in terms of hiding capacity, recovery quality as well as security against steganalysis methods.

Figures (12)

• Figure 1: The Capacity-Distortion Trade-off of hiding different numbers of secret images using different steganography methods, including ISN Lu2021LargecapacityIS, InvMIHNet chen2024InvMIHNet, and SMILENet. The horizontal axis of Capacity-Distortion curve is RMSE of cover/stego image pairs; the vertical axis indicates the hiding capacity of secret images evaluated in terms of the sum of mutual information between input and recovered secret images (see Eqn. \ref{['eq:CD_fun']} for more details). The results are evaluated on DIV2K agustsson2017ntire dataset.
• Figure 2: The hiding capability of the representative image steganography methods in terms of the number of secret images.
• Figure 3: Framework overview of the proposed SMILENet. During the hiding process, the secret images $\{ \bm{x}_{si} \}_{i=1}^N$ are first processed by an SIS module to select essential information $\{ \bm{\tilde{x}}_{si} \}_{i=1}^N$ to be concealed. Next, an ICDM module, guided by the cover image, reversibly transforms $\{ \bm{\tilde{x}}_{si} \}_{i=1}^N$ to secret representations and then spatially splices them to a Mosaic Secret Representation (MSR) $\bm{x}_{ms}$. Subsequently, an IMSE module conceals the MSR into the cover image $\bm{x}_{c}$. Its outputs include a stego image $\bm{x}_{\text{stego}}$ and an image-agnostic component $\bm{r}_{\mathcal{H}}$ which is discarded thereafter. During the recovery process, the reverse pass of the IMSE module and the ICDM module recover MSR $\bm{\hat{x}}_{ms}$ and reconstruct the pre-processed secret images $\{ \bm{\bar{x}}_{si} \}_{i=1}^N$. Finally, a Secret Detail Enhancement (SDE) module refines the recovered images by enhancing their fine details, thereby generating the final secret images $\{ \bm{\hat{x}}_{si} \}_{i=1}^N$.
• Figure 4: The process of splicing and splitting $N=m \times n$ secret images.
• Figure 5: Visualization results of SMILENet on hiding and recovery 25 secret images evaluated on DIV2K, Paris StreetView and ImageNet datasets. All residual images are enlarged 10 times for better perception.