acceptodds
Under review as a conference paper at ICLR 2027

Continuous Detail-Exposure Control via Latent Skipping for View-Level Privacy in Learned Image Compression

Abstract

Learned image compression typically optimizes rate–distortion performance, but offers limited control over how much visual detail is exposed at decoding time. We study controllable detail exposure: suppressing fine-grained texture while preserving a structurally interpretable view, thereby limiting what downstream tasks can extract from the decoded view, a guarantee we refer to as view-level privacy. We propose sCBH (s-Contract with Blockhead), a channel-level skipping framework trained without any downstream-task objective, in which a single scalar (the detail-exposure strength) determines how many latent channels are retained in the bitstream. At , sCBH is bit-exact to the base codec, adding zero quality tax; decreasing progressively removes texture channels from the bitstream and reduces bitrate. A lightweight structural decoder reconstructs a low-detail representation from the structural channels, preserving coarse scene structure under aggressive skipping. Skipping and stuffing strategies are adapted to the causal order of each entropy model. Experiments on ELIC, Hyperprior, CCM, and HPCM show continuous control over bitrate and detail; across four downstream tasks, performance degrades according to each task's reliance on texture versus structure. The source code is publicly available on https://anonymous.4open.science/r/sCBH.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Learned Image Compression, Privacy-Preserving Compression, View-Level Privacy, Continuous Detail Control

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_Nw3XOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript presents a channel-level skipping method for learned image compression. The proposed method sCBH is trained without any downstream-task objective. The claimed contributions include: a single scalar continuously controls the detail-exposure strength, quality, and bitrate; specially designed loss function and the Blockhead, a CNN net for task-agnostic generalization; and a lightweight blockhead decoder attached to the frozen backbone. Experiments are conducted to show continuous control over bitrate and details.

Critical strengths

The paper is well presented. The contributions are well explained. Experiments are clearly presented. The source code is publicly available on anonymous repository (The reviewer has not checked it throughly).

Critical weaknesses

My major concern is on the motivation of this work, i.e., controlling how much visual detail is exposed at decoding time. The intended practical application remains unclear, and the datasets used in the experiments do not align with the assumption underlying the proposed setting.

Questions for authors

What is the intended practical application? Why we need to control the exposed details for decoding? How to decide how much details should be exposed? Who is responsible to control it?

3 comments

We thank the reviewer for the direct assessment. We respectfully disagree with the conclusion that the setting is impractical, and we would like to make the case on three grounds: the deployment exists, the alternatives are worse, and the adoption cost is negligible.

The deployment. The governing principle is data minimization. A receiver should get the minimum detail its purpose requires. An image repository stores one bitstream per image and serves a prefix of it at different levels of detail. The owner keeps the full-quality copy. A viewer whose task is structure-level receives a view sufficient for that task but not for fine detail, which is the part the task does not use and the owner does not wish to release. sCBH is progressive at the channel level. The view at any ss is a prefix of the same channel-coded representation, and the suffix is never entropy-coded. The sender or ingestion service fixes ss before arithmetic coding, so the receiver has no upgrade path. The value follows from a utility floor (DSC, AUC, OCR) for the task and the smallest exposure that meets it. Today the artifact is all-or-nothing, which means the full image, or a degraded copy that cannot be upgraded. sCBH makes the middle ground a property of the artifact itself.

The alternatives do not bound the artifact. Access control decides who receives a copy, not what the copy contains once it is released. The latter is what sCBH controls. Display-time filtering acts on a full-fidelity artifact and is reversible: at matched bitrate Gaussian blur still leaves 98.5% of faces detectable, and post-decoding K-means needs 0.85 bpp / 19.12 dB to reach the zero-detection point that sCBH reaches at 0.154 bpp / 22.17 dB. Pre-processing degrades the stored image permanently and forfeits the full-quality mode. sCBH keeps a single bitstream bit-exact at s=1s{=}1 while making the released view's detail absent rather than filtered.

The cost. sCBH provides a plug-in, not a new codec architecture. It adds a 57K-parameter decoder (vs. 12M for the ELIC decoder) to a frozen backbone, requires no codec retraining, and attaches to ELIC, Hyperprior, CCM, and HPCM with the same recipe. At s=0s{=}0 it is 2.3× faster than ELIC on GPU and at parity on CPU, because only 32 of 320 channels are entropy-coded. A service can add the knob without touching its codec and leave it at s=1s{=}1 until a policy requires otherwise. Even if one does not adopt the privacy framing, this is a low-risk codec feature: 57K parameters, bit-exact at s=1s{=}1, continuous control, validated on four codecs.

The datasets. The four probes instantiate the assumption. Face detection and OCR depend on fine detail, ISIC and CheXpert on coarse structure. The same ss drives face detection from 99.5% to 0% and OCR retention from 86.8% to 8.7%, while ISIC DSC and CheXpert AUC degrade only to 0.539 and 0.733. CelebA is the easiest case, so we add WIDER FACE (3,226 images, 39,123 faces; unaligned, multi-scale, single-scale 640). The same pipeline reduces APeasyAP_{easy} from 0.919 to 0.658 and APhardAP_{hard} from 0.588 to 0.191, with large-face recall at 0.95; the direction holds at λ=0.5\lambda{=}0.5 and λ=0.006\lambda{=}0.006 (Appendix E). Suppression is scale-selective and follows the block size.

What changed. Section 1 states the threat model (decoded view only, fixed non-adaptive models) and points to the residual-identity bound in Appendix C. Sections 4.2 and Appendix E report the WIDER evaluation. We also removed "complete face identity protection", restating the claim as detail/detectability suppression.

We thank the reviewer again for the direct feedback and are happy to clarify any number.

I listed the following simple questions: What is the intended practical application? Why we need to control the exposed details for decoding? How to decide how much details should be exposed? Who is responsible to control it?

The authors did not answer them directly, or manually.

But based on the authors' response to other reviewers' questions, I will raise my rating temporarily.

Thank you for restating the four questions. Our earlier reply was organized by weakness rather than by your questions, so we answer them directly here.

Q1 What is the intended practical application? An image repository, for example, cloud storage that stores one image and releases it at different levels of detail under data minimization. The owner keeps the full-quality copy, while a low-trust viewer, such as an insurance adjuster, a moderation queue, or a non-specialist clinician, receives a view sufficient for coarse content but not for fine detail. Cross-institution medical exchange (organ shape for triage, not skin texture) and document preview (layout, not text) follow the same pattern.

Q2 Why control the exposed detail at decoding time? Access control decides who receives a copy, not what the copy contains once it is released. A display-time filter acts on a full-fidelity artifact and is reversible. As reported in Table 12, Gaussian blur at matched bitrate still leaves 98.5% of faces detectable. sCBH removes the detail from the bitstream itself, so the receiver cannot de-blur or upgrade it.

Q3 How is the amount of detail decided? The sender can set the knob ss to control the amount of detail before the official bitstream is released. The deployer sets a utility floor for the legitimate task (DSC, AUC, OCR) and selects the smallest ss that meets it. The ss-sweep is the calibration instrument and the four probes show what each floor costs in detail.

Q4 Who is responsible for control? The sender or the ingestion service fixes ss before arithmetic coding, so the receiver has no upgrade path to get more details. This is a policy decision at the data controller, implemented by the codec knob. The add-on is 57K parameters and bit-exact at s=1s{=}1, so the deployment cost is negligible.

We hope this addresses each question directly. We are happy to clarify any one of them.

Reviewer_uTo1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes sCBH, which introduces channel skipping, a structural decoder named Blockhead, and inference-stage K-means on top of a pre-frozen learned image compressor.

Critical strengths
  1. the work holds engineering value regarding the verification of actual bitstreams and compatibility-related aspects
  2. the paper is well-structured and logically clear.
Critical weaknesses
  1. The metrics provided in the paper are insufficient to support the conclusions drawn regarding identity privacy. The authors explicitly exclude attack scenarios involving adaptive attacks or the extraction of latent variables. In Appendix A.2, the proportion of samples where both faces are successfully detected is used as the primary privacy metric, effectively equating a detection failure to the impossibility of subsequent identity verification. However, this only demonstrates the failure of the specific "RetinaFace + ArcFace" pipeline employed; it does not prove that identity information cannot be extracted from the decoded images.

  2. The paper fails to clearly distinguish its core advantages from alternative approaches that process images after decoding, while maintaining comparable utility. As shown in Table 6, at a bitrate of 0.1541 bpp, removing the K-means step during the inference phase causes the face detection rate to jump from 0 to 0.8675. This indicates that the reported experimental results are heavily dependent on the post-processing stage. In Table 4, the Gaussian baseline and the configuration with s=0 exhibit similar bitrates but yield different PSNR values ​​(24.52 dB vs. 22.17 dB); moreover, they are not aligned in terms of structural utility or the degree of suppression for sensitive tasks.

Questions for authors
  1. Assuming the attack model is not retrained and only decoded images are used: how does the identity verification performance change if a different, fixed face detection and recognition pipeline is employed (when s is set to a small value)?
  2. For the comparative baselines, if the same post-decoding K-means or blurring operations are applied—ensuring alignment in terms of structural utility and the level of suppression for face/OCR tasks, what are the resulting bitrates, encoding times, and decoding times?
1 comment

We sincerely thank the reviewer for the careful reading. We agree with both points that detection failure does not establish identity infeasibility, and the submitted version over-credited channel skipping for the suppression. The revision corrects the claim and separates the mechanisms with new measurements.

For W1, we have adjusted our claim carefully, removing words such as "complete face identity protection", "identity verification impossible", and the guarantee language. The claim is now detail/detectability suppression under fixed, non-adaptive models. To measure what remains, we align the decoded view with CelebA's 5-point ground-truth landmarks and apply ArcFace directly, removing detection from the loop (Appendix C). This is a worst-case fixed-pipeline stress test, not a deployable attack, and it lies outside the "only decoded view" threat model. On 400 pairs, the standard pipeline detects no faces at s=0s{=}0, yet oracle-aligned ArcFace reaches AUC 0.7585 [0.709, 0.806] and exposure 29.5% at 5% FAR [23.6, 36.2] (11.0% at 1% FAR), against AUC 0.9574 / 92.5% at s=1s{=}1. Exposure at 5% FAR therefore falls from 92.5% to 29.5%, but it is not eliminated. We state this explicitly in the Scope paragraph. A narrower 16-channel variant lowers it further to AUC 0.6720 / 19.0% at 0.078 bpp / 19.10 dB.

For Q1, a second fixed pipeline (an SCRFD detector with a MobileFaceNet recognizer, decoded view only) gives 1.3% detection and 1.5% exposure at s=0s{=}0. The xfullx_{full} side channel, which bypasses the readout and is outside the threat model, exposes 36.0% (32 ch) / 15.0% (16 ch), confirming that suppression comes from the readout rather than from the codec alone.

For W2, i.e. attribution under matched conditions, we agree with the specific observation and we add a rigorous matched comparison. At 0.154 bpp, removing the inference-time readout raises detection from 0 to 86.8%, which is why we now identify K-means as the suppression operator. We separate the roles as follows. Color quantization suppresses details. Channel skipping is the rate mechanism. The 32-channel bottleneck is what makes K-means effective at low rate. The Blockhead is a quality module (Appendix D). At the same K(s=0)K(s{=}0) and the same K-means implementation, applying the readout to a full-rate ELIC decode still detects 32.0% of faces, while the combination detects 0% at 0.154 bpp. At matched suppression (det = 0), compared with ELIC followed by K-means (K=2K{=}2), our sCBH has 5.5× fewer bits and 3.05 dB higher PSNR. At matched rate (0.126 bpp), the baseline still detects 28.0%. The result is therefore not explained by processing location, nor by distortion magnitude: at matched bpp (σ=2.5\sigma{=}2.5) Gaussian detects 98.5% of faces; at matched PSNR (σ=5\sigma{=}5, 22.68 dB, above sCBH's 22.17) it still detects 91.5%. Applied to a full-rate decode it detects 91.0%. A fixed σ\sigma cannot span both ends, whereas sCBH is bit-exact at s=1s{=}1.

For Q2, timing questions, we report encode/decode time including clustering on GPU and CPU (Table 5), showing that our method is practical to deploy. It saves time at low ss on GPU. On an RTX 2080Ti, thanks to channel skipping, s=0s{=}0 takes 307.7 ms versus 720.4 ms for ELIC (2.3× faster); K-means costs 8.6% of the total, so the codec saving dominates. s=1s{=}1 is essentially ELIC on GPU. On CPU the gain narrows to parity because K-means costs 2.32 s, and s=1s{=}1 is 9.3% slower because the Blockhead still runs while K-means is skipped, which can be bypassed in deployment. Post-decoding K-means and Gaussian are slower than sCBH at s=0s{=}0.

What changed. (i) We revised our words in the Abstract, Introduction, and Scope to state the fixed, non-adaptive claim and quantify residual identity information. (ii) Appendix B names the pipeline correctly. (iii) Appendix C adds the upper-bound verification, the 16-channel variant, the second pipeline, and the xfullx_{full} side channel; (iv) Appendix D adds the 2×2 decomposition and matched comparisons; (v) Table 7 and Sec. 4.7 state the three roles.

We hope the revised claim, the measured ceiling, and the matched comparisons address the reviewer's concerns. We are happy to clarify any number.

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

The paper investigates a systematic framework to control how much visual detail is exposed in the learned image compression. In general, the proposed sCBH method skips the selected latent channels and reconstructs a simplified image using a small decoder with K-means post-processing. The detail-exposure strength is controlled by a single scalar s to control the trade-off between detail suppression, distortion, and bitrate. Extensive numerical experiments are conducted.

Critical strengths

The main advantage of this framework is that it is easy to integrate with the existing neural codecs without retraining the compression model. sCBH only trains a small additional decoder and keeps the pretrained compressor backbone.

In addition, the training objective does not require any labels for any task-specific training. Numerical experiments show the framework fits for multiple entropy models, including ELIC, Hyperprior, CCM, and HPCM.

The paper is well written. The presentation is clear and well-organized.

Critical weaknesses

Table 6 shows that removing inference-time K-means increases face detection from 0% to 86.75% at s=0, while bitrate remains unchanged. This indicates that K-means is essential to the reported face suppression and shows the unclear contribution of the chosen latent skipping s for the privacy claim in the paper.

In numerical experiments, Gaussian blur and K-means are evaluated at different settings. For instance, the Gaussian blur baseline has a similar bitrate to sCBH at s=0, but higher PSNR. This makes it difficult to determine whether stronger suppression comes from a better scheme or greater image degradation.

How to show that the proposed method separates the structure and texture information clearly? The early channels can preserve a low-detail image without separating structure and texture information.

Questions for authors

Please see the Weaknesses section.

1 comment

We thank the reviewer for the positive assessment and for three precise points. We have addressed them with matched-distortion measurements and a precise statement of the mechanism. Our conclusions are unchanged, but the division of labor is now explicit.

For what skipping contributes, we agree with the reading of Table 7 (formerly Table 6). At 0.154 bpp, removing the readout raises detection from 0 to 86.8%, and the bitrate is unchanged because the readout is decoder-side. The converse is equally important. Applying the same readout to a full-rate ELIC decode still detects 32.0% of faces, and the best full-rate point that reaches det = 0 is K=2K{=}2, at 0.85 bpp and 19.12 dB. Quantization is the suppression operator, but channel skipping determines the signal it acts on. Neither component alone is sufficient. At s=0s{=}0, the decoder receives only the 32 structural channels and the 32-channel bottleneck is what makes quantization decisive at 0.154 bpp. Skipping also has a representation-level property that blurring does not. The texture channels are absent from the bitstream, not attenuated in the pixels. The full decomposition is in our response to Reviewer uTo1 and in Appendix D.

For the matched comparison, we now compare at three alignments. (i) Matched bitrate: Gaussian σ=2.5\sigma{=}2.5, 0.150 vs. 0.154 bpp, 24.52 vs. 22.17 dB, det 98.5%. (ii) Matched or better PSNR: σ=5.0\sigma{=}5.0, 22.68 vs. 22.17 dB, 0.086 vs. 0.154 bpp, det 91.5%; the sweep remains monotone (σ=6.0\sigma{=}6.0: 82.3%). (iii) Matched operator location: Gaussian applied after a full-rate ELIC decode, 22.74 dB, det 91.0%. The within-method control is decisive: removing K-means raises PSNR by 0.45 dB while detection rises from 0 to 86.8%. A 0.45 dB change is not enough for an 86.8-point change in detectability. The difference is therefore the type of distortion, rather than its magnitude. K-means preprocessing likewise reaches 31.31 dB at 1.02 bpp yet still detects 48.3%.

For the separation of structure and texture, we agree that a low-detail image alone does not establish a separation, and we want to clarify that we do not claim a semantic factorization of the latent. The prefix yielding a low-detail image is the intended effect of channel skipping. What the split specifies is which channels form that prefix and how many are needed. We define it functionally: "structural" denotes the prefix that the codec's own causal order fills first and that still decodes a coherent coarse view; "texture" denotes the remainder, which carries the residual detail. The axis is not specific to one architecture: channel priority in SCCTX, variance order in CCM/hyperprior, step order in HPCM. Fig. 3 shows how we identify it in each case. That the same scalar then produces a similar qualitative behavior across five variants is what we would expect if the axis were a property of learned latents rather than of one model. Three measurements support the choice of 32: the csc_s knee (16 loses 3.1 dB; 48/64 add rate without structural gain), the 22.62 dB reconstruction from the prefix through the Blockhead, and the task asymmetry. The prefix does not remove identity and suppression comes from quantizing that view. Each component is necessary, and the gain comes from composing them. Appendix J shows the two views side by side. Identifying this axis per architecture is also what lets sCBH attach to ELIC, Hyperprior, CCM, and HPCM without retraining the backbone.

We are grateful for the reviewer's positive assessment and for pushing us to separate what the operator does from what the representation enables. We have discussed these changes in Table 7, Sec. 4.7, and Appendix D.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.