acceptodds
Under review as a conference paper at ICLR 2027

StegoCoder: Learning Channel for Steganographic Coding

Abstract

Image steganography hides a message in a cover image with minimal perceptible distortion, producing a stego image from which an authorized receiver can recover the message. Robust steganography additionally requires recovery after the stego image is perturbed. Classical steganographic codes operate close to the rate-distortion bound, but on a fixed binary projection of the cover, typically its least-significant bit planes, in which the payload is destroyed by even mild perturbations. Neural steganography can be trained end-to-end to survive specific perturbations, but lacks the explicit rate-distortion control that coding theory provides. StegoCoder combines the two. A steganographic code first embeds the message into a fixed binary sequence and produces a binary stego sequence. A neural renderer then synthesizes the stego image from the cover image and the resulting binary stego. At the receiver, a learned bit predictor outputs soft estimates of the binary stego from the perturbed image, and the code decodes the message. The networks thus learn a reliable binary channel over which a classical steganographic code operates, splitting the problem into learning the channel and embedding the message efficiently over a binary cover. With a nested polar code, the same code acts as a lossy source code at the sender and as a channel code at the receiver, giving explicit control over the trade-off between embedding distortion and robustness. On CelebA-HQ, StegoCoder recovers most of the test images without a single bit error under additive Gaussian noise of standard deviation at and bits per pixel (bpp) and at and bpp, where least-significant-bit embedding, per-image optimization and end-to-end neural approaches recover none and quantization index modulation substantially fewer. A per-image refinement of the stego image further extends robustness across different payloads and co-training the scheme with the neural polar decoder decreases the distortion.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Data Hiding, Image Steganography, Steganographic Coding

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_cpV2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes StegoCoder, which combines classical steganographic coding (LDGM and polar codes) with a learned image channel. A neural renderer embeds coded binary sequences into images, while a bit predictor recovers them after perturbations. Experiments demonstrate strong robustness to Gaussian noise at high payloads.

Critical strengths
  1. The integration of steganographic coding with a learned image channel is interesting and provides a useful perspective on robust information hiding.

  2. The method achieves strong message recovery under Gaussian noise, particularly at high embedding rates.

Critical weaknesses
  1. There is a classical research direction in image steganography, namely “steganographic distortion cost learning”, which uses neural networks to learn cover-dependent modification cost maps and has a theoretical foundation in rate–distortion theory. However, this relevant line of research is not adequately discussed or cited in the paper.

  2. The framework uses the message and an all-zero binary cover to generate the binary stego sequence. Since the binary cover contains only zeros, the resulting binary stego depends on the message but contains no information from the actual cover image. This is somewhat strange, as the Steganographic Embedding module essentially becomes a cover-independent message encoder, seemingly designed mainly to produce a binary representation for fusion with the cover image. The necessity and advantages of this design are unclear.

  3. The experimental comparisons are limited and mainly involve relatively old methods. Although SteganoGAN and LISO are well-known methods, I am sure they are not competitive baselines in modern image steganography. Therefore, the reported superior performance does not sufficiently demonstrate the effectiveness of the proposed design. Moreover, comparisons with neural steganography equipped with equivalent error-correcting codes are missing, making it difficult to isolate the contribution of the learned channel.

  4. Appendix L reports that 99–100% of the original stego images are flagged by pretrained steganalyzers. The proposed defense is evaluated only against the targeted, frozen detectors, without testing generalization to unseen or adaptively trained steganalyzers. This raises concerns about the practical security of the proposed method.

  5. Figure 1 appears somewhat AI-generated images.

Questions for authors

See weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer cpV2: addresing reviewers concerns

We thank the reviewer for the comments and here we address the questions raised by the reviewer.

  1. We will add this line of work to the Introduction and Section 2.1. Cost-learning methods such as ASDL-GAN (Tang et al., 2017), UT-6HPF-GAN (Yang et al., 2020) and SPAR-RL (Tang et al., 2021) learn the costs ρi\rho_i of Eq. (1) for a fixed binary representation of the cover, such as its LSBs, and embed over a lossless channel. StegoCoder instead learns the representation and the channel (Render and BitPredictor) and uses uniform costs ρi=1\rho_i=1, since it is trained with the same pflipp_{flip} at every position. Combining learned, cover-dependent costs with a position-dependent pflipp_{flip} is a natural extension, which we will state as future work.

  2. The binary stego does not depend on the cover image, but the stego image does: the cover enters through the Render and acts as a channel state known to the sender only (Sec. 3.1). With a fixed binary cover, the code flips only the fraction of positions set by its rate (8.3–12.9%, Table 3), and the image distortion grows approximately linearly with the number of flips (Appendix B), matching the additive distortion assumed by the codes.

  3. We added two coded baselines: QIM with the same nested polar code (response to Reviewer sQJ1, item 2), and the same networks with a standard channel polar code, retrained for its input (response to Reviewer DMrb, item 1). With equivalent codes, both recover (almost) every image at σ=5⋅10−3\sigma=5\cdot10^{-3}. At σ=10−2\sigma=10^{-2}, QIM with the code loses every image at matched PSNR, and the channel-polar networks need LPIPS 0.136 vs. 0.077 for StegoCoder (at 1.3 dB lower PSNR) to recover every image. Regarding recent methods, the comparison includes LISO (2023) and, under JPEG, LISO + neural cover selection (2024).

  4. We tested both (response to Reviewer sQJ1, item 1). SRNet, XuNet and KeNet retrained from scratch on StegoCoder images detect 92–100% of the images defended against the frozen detectors. Rerunning the defence against each retrained detector takes that detector to 0% on all 500 test images, but every other detector, retrained or frozen, still flags almost all of them. The defence therefore evades only the detector whose weights the sender knows. We will add these results in a new appendix, state this in the Limitations, and present StegoCoder as robust data hiding rather than secure steganography.

  5. Figure 1 was drawn by the authors in draw.io, as stated in the Use of Generative AI section. We will redraw it as a plain block diagram (Render, BitPredictor, steganographic code) without decorative elements.

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

StegoCoder combines classical steganographic coding with a learned image channel for robust, high-payload message hiding. At the sender, an LDGM or nested polar construction maps a message and an all-zero virtual cover to a low-Hamming-weight binary sequence. A neural renderer embeds this sequence into a cover image. A bit predictor estimates the sequence after transmission, and the classical code extracts or decodes the message. The networks are initially trained with independent Bernoulli bits, avoiding differentiation through the discrete encoder. The method also uses interleaving, neural-polar-decoder-assisted fine-tuning, and optional pixel refinement; the reported co-trained models use standard SCL decoding at inference. Experiments on CelebA-HQ show strong complete-message recovery at selected 1-4 bpp Gaussian-noise operating points, with additional JPEG, 8-bit quantization, and AFHQ-Dog transfer evaluations.

Critical strengths

S1 A useful interface between coding and learned image embedding. The method separates discrete representation selection from image rendering and bit prediction. The nested polar construction can select low-weight representations at the sender and provide error correction at the receiver, while the networks handle image content and transmission perturbations. Bernoulli-bit pretraining makes this interface practical without backpropagation through the discrete encoder. This is a meaningful system-design contribution, even though its individual coding components are established. S2 Strong complete-message recovery at substantial payloads. In Table 1, standard polar decoding recovers all 500 test messages at 1 bpp under Gaussian noise with standard deviation 0.01. At 2 bpp, the noise-aware-refinement version recovers 456/500 messages under the same noise level. Reporting both BER and complete-message recovery is appropriate for long binary payloads. The 8-bit quantization and AFHQ-Dog transfer tests provide additional evidence that these results are not confined to unquantized, in-domain images. These findings make the proposed interface worth investigating.

Critical weaknesses

W1 The benefit of low-weight shaping over ordinary error protection is not isolated. The method combines shaping, error correction, interleaving, and noise training, but the main comparisons do not isolate the shaping contribution. An informative control is the same learned embedding architecture with a standard channel polar code and no shaping degrees of freedom, appropriately trained and evaluated at matched net payload and comparable image distortion. A neural baseline with standard outer error correction would provide complementary evidence. LDGM differs in both error protection and configuration. Table 3 also reports a 13.6% flip rate for LDGM at 2 bpp, outside its 6-10% training range, further complicating attribution. Consequently, the results establish the usefulness of the full system, but leave the independent advantage of its distinctive coding organization insufficiently demonstrated. W2 The mechanism and independent benefit of decoder-assisted co-training remain unclear. Section 3.3 describes a neural-decoder loss while stating that Render, Fusion, and BitPredictor are fine-tuned with Eq. (4). Does the decoder loss also update these networks? Otherwise, improvements may arise from additional fine-tuning or actual-codeword training. The quality gains accompany large reliability changes: at 2 bpp and Gaussian noise of standard deviation 0.01, the noise-aware-refined standard model recovers 456/500 messages, versus 0/500 for its co-trained counterpart (Table 1). This trade-off can be valid, but a matched fine-tuning control and comparison at equal recovery requirements are needed to establish an improved quality-reliability frontier. An auxiliary training decoder with standard SCL inference is a reasonable possibility; its effective contribution is not yet demonstrated. This concern affects a claimed contribution, beyond presentation alone.

Questions for authors

Q1 Focused validation of the shaping contribution. At one representative payload, can you compare against a properly trained pure channel-polar control? In the 1 bpp configuration, keeping the block length at 1024 and the message length at 257 while removing the 639 shaping degrees of freedom leaves 767 known frozen positions. Please use an appropriate code construction, retrain for its input distribution, count coding overhead, and compare complete-message recovery at matched net payload and comparable PSNR/LPIPS. Q2 Clarification of the co-training objective. What is the complete optimization objective for each module in Section 3.3? Does the neural-decoder loss backpropagate into Render, Fusion, and BitPredictor, or are they updated only by Eq. (4)? Please also clarify the noise distribution used for each module during fine-tuning, given the reported pretraining upper limit of 0.13 and neural-decoder-training upper limit of 0.013. These details determine whether the claimed adaptation mechanism is implemented. Q3 Focused validation of decoder-assisted training. Can you compare decoder-assisted fine-tuning with Eq. (4)-only fine-tuning using the same actual codewords, noise distribution, training budget, and standard SCL evaluation? A small loss-weight sweep, or a comparison at a common complete-message-recovery target, would help determine whether co-training improves the quality-reliability trade-off rather than selecting a different operating point. A positive result on this control could materially improve my assessment of W2.

1 comment
AuthorsOpenReview ↗

Response to Reviewer DMrb: additinal experiments and corrections

We thank the reviewer for interesting view and propositions and here we address the reviewer's questions.

  1. For the proposed control, we ran it at 1 bpp with the same n=1024n=1024, q=257q=257, B=4B=4 planes, interleaver and standard SCL (L=8L=8), so the net payload (65,792 bits) is unchanged. The message occupies the 257 most reliable positions (5G NR order); the other 767 are frozen to a known pseudo-random sequence. A channel codeword has ~50% ones (vs. 8.3% flips for the nested code), so the Render and BitPredictor are retrained with pflip∈[0.45,0.55]p_{\text{flip}}\in[0.45,0.55], sweeping λrec\lambda_{\text{rec}} to match image quality.
code (1 bpp)λrec\lambda_{\text{rec}}PSNR ↑LPIPS ↓0-err, σ=5⋅10−3\sigma=5\cdot10^{-3}0-err, σ=10−2\sigma=10^{-2} (BER)
nested polar (StegoCoder)0.00835.170.077500500 (0)
channel polar, not retrained0.00823.870.37500 (1.2⋅10−21.2\cdot10^{-2})
channel polar, retrained0.00829.470.233500500 (0)
channel polar, retrained0.0231.270.190500500 (0)
channel polar, retrained0.0533.880.136500500 (0)
channel polar, retrained0.136.660.091500325 (4.6⋅10−44.6\cdot10^{-4})
channel polar, retrained0.240.190.0504980 (0.38)

At σ=5⋅10−3\sigma=5\cdot10^{-3} the control has better image quality than StegoCoder, so shaping is not needed there. At σ=10−2\sigma=10^{-2} and comparable PSNR it needs LPIPS 0.136 (500 images) or 0.091 (325) vs. 0.077 for StegoCoder, and at LPIPS 0.050 it recovers none. The advantage of the nested code is thus lower distortion at a common complete-recovery target under σ=10−2\sigma=10^{-2}. We will limit the claims in the abstract and conclusion to σ=10−2\sigma=10^{-2}. The control models were trained for fewer epochs.

  1. We thank the reviewer for pointing this out: the description in Sec. 3.3 is incomplete and partly incorrect. The Render, Fusion and BitPredictor are not fine-tuned with Eq. (4) alone; all modules are trained jointly with one objective:

Lco=λdecLdec+λbitLbit+λrecLrec\mathcal{L}_{co} = \lambda_{dec}\mathcal{L}_{dec} + \lambda_{bit}\mathcal{L}_{bit} + \lambda_{rec}\mathcal{L}_{rec}

with (λdec,λbit,λrec)=(1,0.9,0.05)(\lambda_{dec},\lambda_{bit},\lambda_{rec})=(1,0.9,0.05), where Ldec\mathcal{L}_{dec} is the decoder BCE of Sec. 3.3 (wmsg=5w_{msg}=5), and Lbit\mathcal{L}_{bit}, Lrec\mathcal{L}_{rec} are those of Eq. (4). One Adam optimizer updates the decoder (lr 10−310^{-3}) and the Render, Fusion and BitPredictor (lr 10−410^{-4}), batch 6. Yes, Ldec\mathcal{L}_{dec} backpropagates into all three through the BitPredictor's soft estimates; the polar embedding has no gradient. Noise σ∼U[0,0.0125]\sigma\sim U[0,0.0125] is drawn every step and shared by all modules (0.013 in Sec. 4 is this value rounded).

  1. The comparison below shows that co-training selects a different operating point rather than improving the quality–reliability trade-off. Eq. (4)-only is the same run with λdec=0\lambda_{\text{dec}}=0. All runs use 2 bpp with the same codewords, interleaver and noise; each saved checkpoint is scored with the standard SCL on the first 100 test images. Target: 100/100 at σ=5⋅10−3\sigma=5\cdot10^{-3}.
fine-tuningλdec\lambda_{\text{dec}}λrec\lambda_{\text{rec}}checkpoints meeting targetbest PSNR / LPIPS0-err at σ=7.5⋅10−3\sigma=7.5\cdot10^{-3}
none (pre-trained)–––32.13 / 0.178100
Eq. (4) only00.050 of 12 (max 91/100)––
decoder-assisted0.10.050 of 12 (max 97/100)––
decoder-assisted (Table 1 model)10.0512 of 1234.46 / 0.13153
decoder-assisted, same start10.056 of 934.37 / 0.13251
Eq. (4) only00.025 of 534.59 / 0.12835
Eq. (4) only00.015 of 532.85 / 0.166100

Runs were scored up to 18,047 (λdec=0\lambda_{\text{dec}}=0), 13,861 (λdec=0.1\lambda_{\text{dec}}=0.1), 29,999 (same start) and ~4,300–4,900 steps (λrec=0.01,0.02\lambda_{\text{rec}}=0.01, 0.02). At equal PSNR, Eq. (4)-only (λrec=0.02\lambda_{\text{rec}}=0.02: 34.43 dB, LPIPS 0.131, 53/100) matches decoder-assisted (34.37 dB, 0.132, 51/100). Co-training selects a different operating point but does not improve the trade-off. All fine-tuned models recover at most 18/100 at σ=10−2\sigma=10^{-2} (62/100 before), the trade-off the reviewer noted. We will add this comparison and describe co-training as setting the operating point in the Appendix.

  1. The reviewer is right. The LDGM code flips 13.6% of positions, outside the training range at 2 bpp (0.06–0.10) and 3 bpp (0.06–0.12), so those rows combine missing error correction with an untrained flip rate. We will state this in Table 3 and Section 5.1 and compare the codes at 1 and 4 bpp. The attribution to error correction is supported independently by Appendix E (Table 4), where both codes run on a synthetic bit-flip channel without a network.
Reviewer_sQJ1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes StegoCoder, a image steganography scheme that combines classical steganographic coding with neural networks. StegoCoder first runs an LDGM or nested polar code over an all zero binary cover. A Render network then embeds the binary stego into the image, and a bitPredictor recovers it from the perturbed image. The experiments report error-free recovery on all 500 CelebA-HQ test images at 1 bpp under Gaussian noise.

Critical strengths

The paper runs classical steganographic codes over a learned binary channel. The neural networks learn the channel, while the code handles message embedding. This separation is clear and keeps explicit control over the payload and the number of flips.

The reserved budget of the nested polar code provides error correction that LDGM embedding lacks. The experiments report a clear improvement in error free recovery under noise at similar image quality.

The method also shows recovery ability under controlled Gaussian noise. The experiments report that the polar version recovers all test images at a low payload under stronger Gaussian noise, outperforming the neural and classical baselines in the comparison.

Critical weaknesses

My main concern is that the current experiments do not support the paper's task positioning.

  1. If this is evaluated as classical steganography, Appendix L shows that XuNet and KeNet detect 99.2%–100% of the undefended images. The defense is optimized against these two frozen detectors and evaluated with the same detectors. It therefore does not show security against a retrained steganalyzer.

  2. If the method is instead viewed as robust data hiding or watermarking, the evaluation covers only Gaussian noise with σ≤10⁻² and JPEG models trained separately for q80 and q90. It does not cover common post-processing such as JPEG compression across quality factors, resizing, cropping, blurring or combined attacks. These are standard robustness tests for watermarking.

  3. Table 2 also reports PSNR values of only 18.73–23.99 dB, which raises a clear concern about imperceptibility. The current results establish error correction under a controlled channel. They do not yet show that the method is suitable for real steganography or robust data hiding in an open environment.

  4. In Table 1, StegoCoder uses nested polar error correction, while no error correcting code is described for LISO, SteganoGAN or QIM. The claim that “only StegoCoder keeps the message” therefore mixes the gain from the learned channel with the gain from coding redundancy. I suggest comparing with QIM plus an external error correction code at matched effective payload and image distortion, so that the two contributions can be separated.

Questions for authors
  1. I suggest retraining SRNet or another modern steganalyzer on StegoCoder images and reporting detection rates before and after the sender-side defense. This would show whether the defense works beyond the frozen detectors used during optimization.

  2. I suggest adding QIM with an external error-correcting code at the same effective payload and image quality. A control that keeps the same networks and equivalent redundancy but removes steganographic coding would also help separate the contributions.

3.The paper also contains several typos and unclear descriptions. For example, Line 488 “for for running the experiment workers.” I suggest proofreading the full paper carefully.

4 comments
AuthorsOpenReview ↗

Response to Reviewer sQJ: requested experiments and revised claims

We thank the reviewer for careful consideration of our paper, and here we aim to address the reviewer's concern.

  1. We agree that StegoCoder does not provide security against steganalysis. We retrained SRNet, XuNet and KeNet from scratch on StegoCoder images. They detect 100% of the undefended stego images, and 92–100% of the images defended against the frozen detectors. When we rerun the defence against each retrained detector, that detector flags 0% of the defended images, but every other detector, retrained or frozen, still flags almost all of them. The defence therefore hides the stego image only from the detector whose weights are known to the sender. We will state this in Appendix L and Limitations, as we present StegoCoder as robust data hiding rather than secure steganography.

  2. We agree with reviewers that for proper comparison we should consider QIM with an external error-correcting code. In our experiments, dither modulation (QIM) embeds on all three colour channels, with step Δ\Delta chosen accordingly. The binary cover is the sequence of QIM coset indices of the cover image. The external code is the same nested polar code as StegoCoder: n=1024n=1024, reserved budget nf=128n_f=128, standard SCL with L=8L=8. Its reserved budget provides the error correction. The payload qq is set so that the payload matches StegoCoder: q=341q=341 at 1 and 2 bpp, and q=512q=512 at 3 bpp.

bppmethodPSNR ↑LPIPS ↓0-err, σ=0\sigma=00-err, σ=10−2\sigma=10^{-2} (BER)
1StegoCoder, polar, std. SCL (Table 1)35.330.076500500 (0)
1QIM Δ=24\Delta=24 + polar35.130.1335000 (1.8⋅10−21.8\cdot10^{-2})
1QIM Δ=28\Delta=28 + polar33.810.157500489 (1.5⋅10−51.5\cdot10^{-5})
2StegoCoder, polar, std. SCL (Table 1)32.100.176500287 (4.8⋅10−44.8\cdot10^{-4})
2QIM Δ=32\Delta=32 + polar32.350.1715000 (0.49)
3StegoCoder, polar, std. SCL (Table 1)30.210.2424930 (0.145)
3QIM Δ=36\Delta=36 + polar30.540.2215000 (0.445)

At σ=10−2\sigma=10^{-2}, QIM + polar at matched PSNR loses every image at every payload, while StegoCoder recovers 500 images at 1 bpp and 287 at 2 bpp. At 1 bpp, QIM + polar needs Δ=28\Delta=28 to recover 489 images. That costs 1.5 dB of PSNR and doubles the LPIPS (0.157 vs.\ 0.076). Section 5.1 and the Abstract will state the comparison at σ=10−2\sigma=10^{-2}, and will include these rows. We will also remove the sentence ``Under noise, only StegoCoder keeps the message''.

A control with the same networks but without steganographic coding was also requested by Reviewer 2 (Q1). Please refer to our response to Reviewer DMrb, item 1 (Q1).

  1. We corrected this typo 'for for', as well as other changes after revision, which are highlighted in blue in the revised version.

Additional comments:

  • We agree that the evaluation covers only additive Gaussian noise, JPEG at the quality the model was trained for, and 8-bit quantization. We will position StegoCoder as robust data hiding under these perturbations, and state the restriction in the abstract and the Limitations: ``This work studies the trade-off between embedding distortion and robustness over a learned channel, rather than defence against a broad set of attacks; each model is trained for one payload and perturbation, and resizing, cropping, blurring, other JPEG qualities and combined attacks are left for future work.''
  • Regarding the low PSNR under JPEG (Table 2): at 1 bpp, JPEG quantization removes the detail that carries the payload, so the methods that report results under the actual encoder operate at a similar PSNR: LISO-JPEG and LISO + NCS report 19.72 and 21.53 dB at quality 80 (BER only), against 18.73–20.36 dB for StegoCoder, which also recovers up to 329 of 500 images without a bit error.

Thanks for the revision and running both experiments. The QIM + polar comparison answers my worry about the coding gain. But the steganalysis result is the problem. The response calls the paper robust data hiding, but the abstract says robust steganography and never mentioned any robust data hiding before the Limitations. Please use robust data hiding in the abstract and the main text. With those fixed I will raise my rating to 3.

AuthorsOpenReview ↗

A follow up for Reviewer sQJ1 with corresponding text changes

Thank you for the follow-up! We agree the framing was inconsistent and have revised it throughout the paper (changes in blue). The Abstract now opens with robust data hiding and says that undetectability by steganalysis is not a goal. The Introduction distinguishes secure steganography from robust data hiding and places StegoCoder in the latter. The Fig. 1 caption, Sec. 2.1, Sec. 3, the Conclusion and the Ethics statement use the same framing. The Limitations and Appendix L now refer back to it instead of introducing it. Though we kept “steganographic coding” in the title as it names the coding technique StegoCoder uses.

Thanks for taking the feedback seriously. So I'm raising my rating to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.