acceptodds
Under review as a conference paper at ICLR 2027

LW-MIH: GENERATIVE MULTI-IMAGE HIDING WITH VAE PACKING AND LEARNABLE WHITENING

Abstract

Multi-image hiding conceals several secret images in one image sent over a public channel, but existing methods face a critical issue: cover-based methods reach high capacity yet modify the cover and leave traces, whereas coverless generative methods are secure yet hide only a few images, and neither states how many images one cover can carry nor how to trade that number against per-image fidelity. To address this issue, we propose LW-MIH, a generative coverless multi-image hiding framework with high capacity and high recovery fidelity. LW-MIH first encodes every secret with a frozen foundation VAE and packs the latents into a fixed-size tensor, so the number of images one cover carries follows from the encoder geometry: up to 96 with Sana, 48 with SD1.5 and 12 with FLUX. It then fills the tensor in one of three ways — full capacity, high-frequency dual-slot and upsampling — turning the number of hidden images and the fidelity of each into a discrete set (FLUX 12/6/3, SD1.5 48/24/12/3, Sana 96/48/24/6). Finally, an invertible neural network maps the packed tensor to the noise end of a frozen rectified flow, and a learnable invertible whitener aligns the field with the flow's Gaussian prior, adding no capacity cost and keeping recovery exact. Experimental results show that at full capacity the 16-bit PSNR of the recovered secrets is 35.03, 28.17 and 26.89 dB for FLUX, SD1.5 and Sana, within 0.15, 0.07 and 0.06 dB of their VAE ceilings, and rises to 41.03, 37.08 and 37.91 dB when fewer images are hidden; the 8-bit values are 24.12, 22.77 and 20.18 dB. The whitener gains 3.97 and 3.36 dB of 8-bit PSNR over a random orthogonal mixer of the same form on FLUX and SD1.5, and at full capacity leaves a worst gap of 0.41 dB where the pipeline without a mixer falls 12.4 dB below its encoder's limit.

Primary area:
generative models
Keywords:
multi-image hiding, learnable whitening, generative steganography, VAE latent packing, invertible neural network, rectified flow

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_4MFwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LW-MIH is proposed as a VAE-based generative framework that addresses all three questions. FLUX, SD1.5, and Sana are used only as frozen secret encoders, while a separate rectified flow trained on the AFHQ-Cat dataset is employed to synthesize the cover. The encoder latents are packed into a single tensor of fixed size, the tensor is filled under three hiding schemes, and the packed field is aligned with the flow prior through a learnable invertible whitener (LIW). Up to 96 images are hidden in one cover without the widespread failures observed in the untrained pipeline.

Critical strengths

This paper derives a capacity formula based on VAE geometry and shows that the maximum number of hidden images is determined by the tensor size together with the downsampling and channel parameters of the VAE. A Learnable Invertible Whitener (LIW) is further proposed to calibrate the amplitude and heavy-tailed distribution and to suppress extremely large values. The ablation studies are well designed and involve rigorous control of variables.

Critical weaknesses

(1) The proposed method relies on an additional information σ\sigma for recovery, and reconstruction performance degrades severely if σ\sigma suffers bit errors, loss, or tampering. Only simple tests are conducted in the appendix. An analysis of this single-point vulnerability should be provided in the main text, together with fault-tolerant schemes.

(2) A cryptographic security analysis is missing. The key space and brute-force complexity are not evaluated, and security under attack models such as ciphertext-only attacks and chosen-secret attacks is not discussed.

(3) The Patch-CNN experiment shows that the cover images are not statistically undetectable. Beyond Patch-CNN, more mainstream steganalysis tools are not adopted for detection experiments. It should be clarified how the detection resistance of cover images can be improved.

(4) All experiments are conducted under an ideal lossless channel, and recovery performance under real-world channel distortions, including JPEG compression, noise, cropping, and scaling, is not tested. Since channel robustness is a critical metric for steganography, relevant experiments should be supplemented, or the lossless-channel assumption should be stated explicitly.

(5) The three packing strategies (FC/HFD/UP) perform allocation only by packing VAE latents; the underlying bit-capacity upper bound is not raised, and bits are merely redistributed among different secret images. The advantages and disadvantages of this design relative to latent compression schemes should be elaborated.

(6)The HFD strategy provides little benefit and slightly decreases PSNR with the FLUX encoder, and positive benefits are observed only for Sana. Although HFD is presented as a general-purpose scheme, it is compatible with only a subset of VAEs, and the underlying reason for this behavior remains unanalyzed.

Questions for authors

(1) The proposed method relies on an additional information σ\sigma for recovery, and reconstruction performance degrades severely if σ\sigma suffers bit errors, loss, or tampering. Only simple tests are conducted in the appendix. An analysis of this single-point vulnerability should be provided in the main text, together with fault-tolerant schemes.

(2) A cryptographic security analysis is missing. The key space and brute-force complexity are not evaluated, and security under attack models such as ciphertext-only attacks and chosen-secret attacks is not discussed.

(3) The Patch-CNN experiment shows that the cover images are not statistically undetectable. Beyond Patch-CNN, more mainstream steganalysis tools are not adopted for detection experiments. It should be clarified how the detection resistance of cover images can be improved.

(4) All experiments are conducted under an ideal lossless channel, and recovery performance under real-world channel distortions, including JPEG compression, noise, cropping, and scaling, is not tested. Since channel robustness is a critical metric for steganography, relevant experiments should be supplemented, or the lossless-channel assumption should be stated explicitly.

(5) The three packing strategies (FC/HFD/UP) perform allocation only by packing VAE latents; the underlying bit-capacity upper bound is not raised, and bits are merely redistributed among different secret images. The advantages and disadvantages of this design relative to latent compression schemes should be elaborated.

(6)The HFD strategy provides little benefit and slightly decreases PSNR with the FLUX encoder, and positive benefits are observed only for Sana. Although HFD is presented as a general-purpose scheme, it is compatible with only a subset of VAEs, and the underlying reason for this behavior remains unanalyzed.

(7) Both encryption and decryption require forward/backward ODE integration of the rectified flow, which leads to prohibitive inference cost. Quantitative reports of inference time and GPU memory consumption are lacking.

Reviewer_29NZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LW-MIH, a generative multi-image hiding framework. It uses frozen VAEs to encode secret images, packs their latents into a fixed-size tensor, and applies an INN and a learnable invertible whitener (LIW) before generating the cover with a rectified flow. The framework includes three hiding schemes and evaluates FLUX, SD1.5, and Sana, reporting up to 96 secret images in one cover.

Critical strengths
  1. The framework supports different VAE encoders and offers several operating points for trading the number of hidden images against recovery quality.
  2. The reported 16-bit recovery results are close to the corresponding VAE reconstruction references. LIW also reduces severe recovery failures in the evaluated FLUX and Sana groups.
Critical weaknesses

1.The paper says the INN only matches marginal distributions, but Eq. (3) models the joint density, so the reason for adding LIW needs a clearer explanation. 2.Since the reported statistics are measured before normalization, I cannot tell whether the generator actually receives a better input. 3.It is hard to tell whether the gains come from learning the mixing matrix or simply adjusting the scale, as these two changes are not tested separately.

Questions for authors
  1. The author says that INN only matches the marginal distributions, but Eq. (3) models the joint density. Local correlations remaining may be due to limited model capacity or insufficient training. This might raise doubts about whether the LIW module is necessary.
  2. Table 11 and Figure 5 report the statistics for ww before normalization, but the generator receives normalized ζ=w/σ\zeta=w/\sigma. A lower raw standard deviation and lower tail mass do not necessarily mean better input to the generator. These changes alone also do not explain why recovery quality improves.
  3. Why is γ=0.8\gamma=0.8 used to transform only some of the coordinates while leaving the rest unchanged? The reason for this choice is not clearly explained. In Table 10, it is difficult to tell whether the improvement comes from the parameters or simply from those four samples being easier. The author should use the same test groups and include a full-mixing control with γ=1\gamma=1 to check whether retaining the identity part is necessary.
  4. Table 8 quotes results from other papers as background, but the experimental settings differ substantially. These numbers do not allow a meaningful comparison of LW-MIH with existing methods.
  5. The current security evaluation uses a very small cover set and a simple Patch-CNN. The author should consider to evaluate on a larger independent test set using established steganalysis networks, such as SiaStegNet or XuNet.
  6. The terms “INN only” in the main text, “Joint INN only” in Figure 3, and “Joint training” in Table 13 are not consistent. Please clarify how each baseline was trained. Also, the architecture and training details of the INN’s coupling sub-networks also need to be provided.
Reviewer_HwHeOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LW-MIH packs multiple secret images encoded by frozen VAEs into a fixed tensor, generates a carrier through an INN, a learnable invertible whitener (LIW), and a frozen rectified flow, and reverses the pipeline for recovery. Three VAEs support up to 12, 48, and 96 images, with additional capacity–fidelity operating points. Recovery approaches the VAE reference at 16-bit precision. LIW improves 8-bit recovery in some configurations, but its mechanism, competitiveness, and threat model require clarification.

Critical strengths
  1. The paired ablation is informative about both the effects and limitations of different mixing strategies. On FLUX, the 8-bit gain over random mixing is substantial. On SD1.5, the same contrast must be read with care: random mixing is itself worse than no mixing, so the +3.36 dB largely restores degradation introduced by that baseline, and LIW ends at 22.77 dB versus 22.88 dB without a mixer. These paired results are informative while making the baseline dependence of the improvements explicit.
  2. The cross-encoder implementation under a fixed tensor budget is clear. Three frozen VAEs, multiple secret-image datasets, and several packing schemes demonstrate implementation flexibility. The authors also disclose the auxiliary scalar, substantial differences between bit depths, and the statistically inconclusive improvement over random mixing on Sana, helping readers assess the method’s scope.
Critical weaknesses
  1. The mechanism explanation lacks key evidence. Equation (3) is a full change-of-variables likelihood for fitting the joint distribution; its coordinate-wise Gaussian term does not imply marginal-only matching. Table 11 reports pre-normalization statistics, whereas the generator receives w/σ, so reduced tail mass may partly reflect scaling. Invertibility also does not establish quantization stability. Post-normalization diagnostics and separate scaling/orthogonal-learning ablations are needed.
  2. The overall system advantage remains insufficiently established. The 196,608 elements are latent scalars, so image-count advantages require rate comparisons at matched distortion. Evaluation groups are few, JPEG/resizing/cropping results are absent, and near-VAE recovery primarily holds at 16-bit precision. On Table 13’s 24 carriers, LIW has worse FID and detection deviation than random mixing. Moreover, public constants replace per-carrier σ with less than 0.02 dB change in mean 8-bit PSNR, leaving payload confidentiality and the receiver-held secret unclear.
Questions for authors
  1. Could you report correlations and tail statistics after w/σ normalization and separately ablate orthogonal learning and diagonal scaling? Does “no recovery-side gain” imply no amplification of quantization perturbations?
  2. Is LIW shared across hiding schemes or trained per VAE–scheme pair? Please clarify data splits and the composition and independence of each statistical “group”.
  3. Could you provide comparisons matching bit depth, auxiliary information, and reconstruction distortion, and clarify the supported channel conditions? How do you explain improved recovery alongside worse carrier metrics in Table 13?
  4. Given near-equivalent average recovery using public constants, what secret does the legitimate receiver possess that an observer does not? If only communication concealment is intended, please limit confidentiality claims accordingly.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.