acceptodds
Under review as a conference paper at ICLR 2027

STAD: Transferable Active Defense against Generative Image Steganography via Structural Feature Perturbations

Abstract

Generative image steganography (GIS) directly generates stego images from secret information, making conventional passive defense methods that rely on detecting modification traces difficult to apply. Most existing active defense methods disrupt covert communication through image degradation followed by reconstruction. However, these approaches are ineffective against GIS, where the carrier image is generated directly rather than modified to embed secret information. We propose Structure-guided Transferable Active Defense (STAD), which perturbs stego images using a single frozen encoder. Neither the hidden message nor the target decoder is required. STAD optimizes differences between neighboring features at several encoder layers, avoiding a loss tied to a specific message representation. It combines gradients from nearby images with edge weighting and multiscale smoothing, then constrains distortion in the quantized output. Tests on five SOTA GIS methods give the lowest bit extraction accuracy among the evaluated defenses and an average 8-bit message accuracy of 17.06%. The surrogate and optimization settings remain fixed across targets, and output PSNR stays at or above 30 dB. Experimental results demonstrate that the proposed STAD effectively defends against GIS methods while maintaining high visual quality.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Generative Image Steganography, Active Steganography Defense, Cross-Architecture Adversarial Transferability, Structural Feature Perturbation, Message Recovery Disruption.

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

This paper proposes STAD, a structure-guided transferable active defense that protects GIS generated by GANs and diffusion models without access to the secret message, decoder key, or target decoder. A frozen VQGAN encoder is adopted as a fixed surrogate, the horizontal and vertical spatial relationships of its intermediate features are perturbed, and the optimization is regularized by neighborhood aggregation, edge weighting, and multiscale smoothing under a post-quantization distortion constraint; the encoder and defense hyperparameters are fixed across all tested GIS targets, and only image-wise perturbations are optimized per input image.

Critical strengths

This paper addresses the key limitation that most existing active defenses are designed for traditional modification-based steganography and exhibit poor transferability against GIS constructed by GANs and diffusion models. To address this issue, STAD is proposed as a black-box defense that uses a single frozen surrogate model with fixed hyperparameters to counter both categories of GIS under settings that closely reflect real-world attack--defense scenarios; experimental results indicate its effectiveness.

Critical weaknesses

(1) GIS scenarios with embedded error-correcting codes (ECC) are not evaluated. Since ECC can recover a certain proportion of bit errors, the robustness of the proposed method under ECC remains unclear.

(2) STAD assumes that secret information is encoded within local spatial neighborhood correlations in deep VQGAN features. It remains unclear whether STAD would fail for newly designed GIS schemes that do not rely on such local spatial feature structures.

(3) Perturbations are added to images by STAD; however, corresponding steganalysis experiments on these perturbations are not reported. It is therefore unclear how it can be guaranteed that the proposed method will not be identified by third parties as steganographic images.

(4) The loss function, gradient sampling, edge weighting, smoothing, and distortion constraints are involved, but the purpose of each introduced component is not clearly explained, and targeted ablation experiments for each module are lacking.

(5) A frozen VQGAN encoder is adopted directly as the surrogate model, and defense performance with other pre-trained encoders (e.g., CLIP, DINOv2, ResNet) as surrogates is not compared.

Questions for authors

(1) GIS scenarios with embedded error-correcting codes (ECC) are not evaluated. Since ECC can recover a certain proportion of bit errors, the robustness of the proposed method under ECC remains unclear.

(2) STAD assumes that secret information is encoded within local spatial neighborhood correlations in deep VQGAN features. It remains unclear whether STAD would fail for newly designed GIS schemes that do not rely on such local spatial feature structures.

(3) Perturbations are added to images by STAD; however, corresponding steganalysis experiments on these perturbations are not reported. It is therefore unclear how it can be guaranteed that the proposed method will not be identified by third parties as steganographic images.

(4) The loss function, gradient sampling, edge weighting, smoothing, and distortion constraints are involved, but the purpose of each introduced component is not clearly explained, and targeted ablation experiments for each module are lacking.

(5) A frozen VQGAN encoder is adopted directly as the surrogate model, and defense performance with other pre-trained encoders (e.g., CLIP, DINOv2, ResNet) as surrogates is not compared.

(6) The efficiency of the proposed model is not analyzed. In particular, quantitative metrics such as FLOPs, per-image inference latency, GPU memory consumption, and throughput are absent from the evaluation.

Reviewer_T9JvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes STAD, an active defense method for generative image steganography (GIS). It employs a frozen VQGAN encoder as a surrogate feature extractor, generates image perturbations by perturbing the spatial differential relationships of multi-layer features, and combines neighborhood gradient aggregation, edge weighting, multiscale smoothing, and output distortion constraints.

Critical strengths
  1. The proposed STAD is interesting and offers a promising direction for research into active defense against GIS attacks.
  2. The experimental results also show that it can reduce the extraction accuracy of various GIS methods.
Critical weaknesses
  1. The paper points out that existing active defense methods based on image degradation and reconstruction are ineffective against GIS. However, it does not fully demonstrate the basis for this conclusion. Although GIS images are generated directly through the generation process, the receiving end may still perform denoising, reconstruction, or generative restoration on the input images before extracting the message. If these operations can eliminate the perturbations introduced by STAD while retaining or restoring message-related information, the effectiveness of current defenses may be significantly reduced. Although the paper tested JPEG compression, scaling, blurring, median filtering, and noise addition, these transformations do not account for reconstruction or generative restoration attacks. It is recommended to supplement the testing with representative image reconstruction, generative restoration, especially diffusion model regeneration tests, and to clarify which processing steps the defender and the receiver can respectively observe and control.
  2. The paper uses the frozen encoder of an ImageNet-pretrained VQGAN as the surrogate but does not explain why its features and spatial differences exhibit a stable correlation with the message encoding mechanisms in various GIS methods. GANs and diffusion-based GIS may convey messages through different latent variables, noise, conditional signals, or generative mappings, respectively. The reason why perturbations to VQGAN feature differences can universally disrupt these decoding processes is currently supported primarily by results, with a lack of mechanistic analysis.
  3. The paper describes the selection of hyperparameters on the RDI-SWE development set, while RDI-SWE is also one of the five primary evaluation targets. Although the authors note that decoder feedback is used only during the development phase and not for generating perturbations to specific images, it is still necessary to clarify the strict separation between the development and test sets, whether the results extracted by the target method were used in parameter selection, and whether the testing of RDI-SWE was independent of the hyperparameter tuning process. It is recommend adding a GIS objective that is completely independent of proxy selection and hyperparameter selection, or conducting a cross-objective leave-one-out test, to more strongly demonstrate that STAD does not rely on information from specific objective methods.
Questions for authors
  1. Does regenerate attacks weaken STAD's defensive effects?
  2. Why can a frozen VQGAN encoder serve as a general-purpose agent across GIS methods?
  3. Did the parameter selection process affect the test results for RDI-SWE?
Reviewer_tbC5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes STAD, an active defense against generative image steganography. STAD generates image-specific perturbations using a frozen VQGAN encoder, without accessing the target decoder or hidden message during perturbation generation. Its objective maximizes discrepancies between spatial differences in intermediate features. The optimization combines neighborhood gradient sampling with edge weighting, multiscale smoothing, and momentum, followed by a distortion check on the quantized output.

Critical strengths
  1. The problem setting is meaningful, and the reported transfer across targets is promising. Generating defensive perturbations without querying the target decoder addresses a practical limitation of task-specific attacks. Using the same surrogate and fixed settings across five targets is a useful evaluation choice. The reported average EA improves from 81.95% for EBRA, the strongest listed baseline, to 72.84% for STAD, while average PSNR increases from 25.34 to 30.35 dB.

  2. The proposed objective has some empirical support, and output distortion is checked after quantization. On RDI-SWE, replacing the spatial-difference objective with the global feature loss increases EA from 76.73% to 88.05%; removing smoothing increases it to 95.88%. These results suggest that the objective and gradient processing matter. Evaluating extraction from the saved 8-bit images also makes the distortion and extraction measurements more representative of the actual output.

Critical weaknesses
  1. The appendix should be directly attached to the main manuscript. The method combines several components, but the ablations do not clearly show which are necessary. In particular, edge weighting and ILPD are not tested separately, and the benefit of gradient-guided sampling over ordinary random sampling remains unclear. Figure 4 also shows that fewer iterations can outperform the default setting, which raises questions about whether the added complexity is justified.

  2. The experiments use only a VQGAN surrogate encoder, so it is unclear how sensitive the results are to this choice. Testing an alternative encoder would help establish whether the observed transfer comes from the proposed objective or depends heavily on the particular surrogate.

  3. Section 4.1 defines EA as the fraction of correctly recovered bits and MA as the fraction of consecutive 8-bit segments recovered entirely correctly. When both metrics use the same complete segments and handle decoding failures consistently, they must satisfy MA≥max⁡{0,  1−8(1−EA)}\mathrm{MA}\geq \max\{0,\;1-8(1-\mathrm{EA})\}, with accuracies expressed as fractions. Every incorrect segment requires at least one incorrect bit, so the number of incorrect segments cannot exceed the number of incorrect bits. This bound does not assume independent bit errors and is preserved by averaging per-image accuracies.

Several LDStega entries in Table 1 appear to violate this relationship. Excluding incomplete trailing segments can change the precise bound, but that convention alone does not explain all of these discrepancies. The STAD entries themselves do not violate this bound. Nevertheless, the discrepancies raise concerns about the shared evaluation procedure and the reported comparisons in MA. This issue needs to be resolved before the main quantitative conclusions can be assessed reliably.

The paper is overall good. I am willing to raise my score if the concerns are addressed.

Questions for authors

See weakness.

Comments on the paper

1 comment
Area_Chair_hp7POpenReview ↗

Please Participate in the Discussion Phase

Dear Reviewers,

The authors have posted their rebuttal and responses addressing your review comments and questions.

Please carefully read their responses (as well as the updated manuscript, if applicable) and actively engage in the reviewer-author discussion phase. If you have any remaining concerns or follow-up questions, please post them so the authors can address them before the discussion period closes.

Your timely participation and updated assessments are essential to helping us reach a fair and well-informed recommendation for this submission.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.