Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Joint Lossless Compression and Steganography for Medical Images via Large Language Models

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read By splitting medical images into significant and insignificant bit planes, a joint lossless-compression and steganography framework embeds patient metadata invisibly while reporting the lowest bits-per-pixel on seven medical datasets.

desk verdict The framework is a real engineering attempt, but its headline compression claim is unmeasurable because the secret-message cost is not separated from the compression bitrate. read the letter →

arxiv 2508.01782 v4 pith:BP3HZPQD submitted 2025-08-03 eess.IV cs.CV

classification eess.IVcs.CV
keywords losslessimagecompressionmedicalsteganographylargelanguagemodelsbitplaneslicingdual-pathreversibledatahidingentropycodingA-LoRA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that medical image compression and patient-metadata hiding can be done in one lossless operation, and that doing so beats existing codecs in bitrate. The method first splits an image by bit planes into a high-significance global part and a low-significance local part. The global part is compressed with a variational autoencoder using bits-back coding, while the local part is compressed by an LLM after the secret message has been spliced into its bit planes. The authors report the lowest bits-per-pixel on seven medical datasets, for example 2.25 bpp versus 2.34 for the DLPR baseline on the chest X-ray set, while the embedded message remains visually and statistically hidden. If the comparison is fair, the practical payoff is a single bitstream that is both highly compressed and carries identity data with no parseable plaintext metadata header.

What carries the argument

The load-bearing object is the adaptive bit-plane decomposition indexed by $s$: the original image $x$ is written as $x = \sum_{l=1}^{m} 2^{l-1} x_l$, and $s$ is the largest index such that $\sum_{i=1}^{s} I(x_i;x) \le (1-\beta) I(x;x)$. This single split determines what goes into the VAE path (global planes $x_{s+1:m}$), what can be overwritten by the message (local planes $x_{1:s}$), and what the LLM must model. The steganographic mechanism is the segmented replacement $Y_i = x_i \dot{R}_{L_i} M_{L_i}$, with recovery via $B_i = x_i \oplus Y_i$ and the transmitted lengths $L_i$; the compression mechanism is arithmetic coding under the binary probability $p_t^{(k)}(v)$ produced by the LLM from a task prompt, visual-prompt embeddings, and previously encoded binary symbols. A-LoRA changes only the LoRA initialization, setting the adapter's low-rank matrix to the mean and variance of anatomical features extracted by a lightweight pretrained network.

What would settle it

Use the released code to compress C-19-R images with a patient record of a realistic length, say 256 to 1024 bytes, and recompute the total bits per pixel including the embedded message and all side information; if the total is not below 2.34 bpp, the claimed improvement over DLPR depends on an unstated tiny payload.

Watch

Extended reading notes

Core claim

The central claim is that an explicit bit-plane split creates a natural place to hide reversible metadata without hurting compression. Adaptive modalities decomposition picks the largest slicing index $s$ such that the low planes contribute at most $(1-\beta)$ of the image's total mutual information; those planes are declared the local modality. The message $M$ is split into segments, each segment replaces one local bit plane's bits, and the XOR difference bitmap $B$ is stored as side information so both image and message can be exactly reconstructed. The stego local planes are then compressed patch-wise by an LLM whose next-symbol probabilities drive arithmetic coding, with global-modality embeddings supplied as visual prompts, and a fine-tuning strategy called A-LoRA initializes the adapter from anatomical-feature statistics. The paper reports that this pipeline reaches 2.25 bpp on the chest X-ray set (DLPR: 2.34), keeps PSNR around 68.79 dB while changing only 0.3521% of pixels, and lowers the SR-Net steganalyzer AUC to about 0.73 at 0.2 bpp payload versus 0.83 for the sequential baseline and 0.92 for vanilla LSB.

Load-bearing premise

The headline bitrate comparison assumes the hidden message is small enough that the extra bits it adds, plus the recovery map that must be sent along, do not erase the reported gains; the paper never states how long the embedded message was in the compression experiments.

Editorial extensions

If this is right

  • On all seven tested medical datasets, the full bitstream—compressed image plus message plus side information—is reported to be smaller than the outputs of classical codecs, learned codecs, and the reproduced LLM baseline, with the largest margins on head CT (1.54 bpp) and lung CT (1.83 bpp).
  • An authorized receiver who decodes the bitstream can reconstruct both the original image exactly and the patient-identity message, because the bitmap $B$, segment lengths $\{L_i\}$, and location index $i$ are all included in the transmitted stream.
  • Splitting the payload across several low-significance planes lowers steganalytic detectability compared with both vanilla LSB and continuous sequential embedding, while keeping the scheme fully reversible.
  • Restricting the LLM to the local modality cuts decoding time from 287.30 s for the pure LLM codec to 39.91 s for the 1.5B variant on the same dataset, while improving bpp from 3.02 to 2.25.
  • LLM size matters only mildly: the 1.5B, 3B, and 7B variants land at 2.25, 2.27, and 2.25 bpp respectively, so smaller models can be deployed with acceptable loss.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the Table II bitrate includes the embedded message and side information in Eq. (19), while baseline bpp counts only image bits; the paper never states the message length $L$ used in the compression tables, so the claimed 4.0% gain over DLPR on the chest X-ray set could shrink or reverse if $L$ were as large as the 0.2 bpp payload tested in the security experiments.
  • Beyond the paper: because the bitmap $B$ is computed from the XOR of original and stego planes, it is an image-specific recovery map that must be transmitted; a natural extension would be to derive $B$ from a secret key at decode time, turning the scheme into blind reversible embedding at the price of some additional side information.
  • Beyond the paper: the same global/local bit-plane split could be applied to other high-bit-depth medical modalities such as volumetric DICOM series or whole-slide images, where the decomposition might be defined over slices or tiles instead of pixel bit planes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a joint lossless compression and steganography framework for medical images. It decomposes an image via bit-plane slicing into global and local modalities, compresses the global modality with a VAE/bits-back codec and the local modality with an LLM-based arithmetic coder, embeds a secret message into multiple local bit planes in segments, and uses an anatomical-priors LoRA fine-tuning strategy. The authors report state-of-the-art bpp results on seven medical datasets and provide security evaluations using SR-Net, JS divergence, and PSNR.

Significance. If the results hold, the framework would be a meaningful contribution: it is one of the first attempts to combine LLM-based lossless compression with reversible steganography for medical images, and it addresses an important practical need for secure transmission of patient metadata. The paper has clear strengths: evaluation across seven datasets of different modalities, runtime and throughput analyses relative to LLM-based codecs, component ablations, and an explicit description of the transmitted bitstream composition in Eq. (19). However, the headline compression comparison is not currently measurable because the message length L is unreported and the baseline codecs transmit only the image, while the proposed bitstream also carries the secret message and side information. In addition, the component ablation contains an information-theoretically implausible result that adding steganography lowers the bitrate. These issues are load-bearing for the central SOTA-while-secure claim and need to be resolved before the paper can be accepted.

major comments (3)
  1. [IV-B1, Eq. (19), Table II] The central SOTA claim is not measurable as stated. Eq. (19) and the Table II caption state that the reported bpp for 'Ours' includes S_L, S_B, S_i, and the compressed stego local planes, while baseline codecs transmit only the image. The message length L is never reported for the compression experiments. On C-19-R (300×300), the claimed margin over DLPR is 0.09 bpp, i.e., 8,100 bits. The security experiments in Table VII sweep payloads up to 0.2 bpp (18,000 bits), and even at 0.1 bpp (9,000 bits) the bitmap B alone carries at least L bits of information before counting S_L, S_i, or the possible degradation of arithmetic coding on the stego planes. The paper therefore leaves open two readings: either a negligible payload was used, making the steganographic component untested in the headline comparison, or a meaningful payload was used, making the comparison unfair. The authors must report L used in Table II and rerun the comparison under matched information content, or explicitly state that the baselines should also transmit the same metadata.
  2. [IV-C1, Table VIII] The ablation reports that adding steganography reduces bpp (Dual-Path 4.09, +Vanilla LSB 3.77, +Segmented Steg. 3.34). This is information-theoretically implausible for random, incompressible message bits: replacing bits of x_{1:s} with M and then transmitting the stego planes together with B, L_i, and i cannot cost less than transmitting x_{1:s} alone, unless the message or side information is not actually included in the reported bpp, or the 'randomly but differently generated' messages have special structure. Please clarify the bit accounting and verify with a concrete example that the stego bitstream plus S_B indeed reconstructs both image and message at the reported bitrate.
  3. [IV-A2 and IV-C2-4] Hyperparameters β, LoRA rank/alpha, and patch size are selected by ablations on C-19-R, and the resulting configuration is then reported as the SOTA result on that same dataset in Table II. This is selection on the test set for the headline dataset. The other six datasets are not used for these ablations and provide some independent support, but the C-19-R SOTA claim should be treated as overfit unless a validation split is used or the authors demonstrate that the chosen hyperparameters are stable across all datasets. Please re-run the ablations on a held-out split or temper the C-19-R-specific claim.
minor comments (6)
  1. [Table II] The VVC-Intra row is identical to the L-Infinite row (2.73, 4.80, 4.02, 3.52, 4.71, 3.09, 3.20), which is likely a copy-paste error; please verify the VVC-Intra results.
  2. [Tables II and IV] Please indicate which LLM size (1.5B, 3B, or 7B) corresponds to the 'Ours' row in Table II and in the main text; Table IV lists three variants but Table II does not specify.
  3. [Figure 4] The caption says that L_1, L_2, L_s, and Bitmap B are sent to the receiver, but the text and Eq. (19) also include the location index i; please update the caption for consistency.
  4. [III-D] The initialization notation A=N(Y_{1:s}; μ_f, σ^2_f) is dimensionally unclear because A is a low-rank matrix while Y_{1:s} is a tensor; please specify how the anatomical feature statistics are mapped onto the LoRA initialization.
  5. [Figure 5] The figure reports 'Total pixels: 82,928' for what appears to be a 300×300 C-19-R image (90,000 pixels); please clarify whether the image was cropped or whether this is a typo.
  6. [IV-B3] The claim that the method achieves 'remarkable invisibility to human observers' is based on PSNR/SSIM and a single visualization; please soften the wording or support it with a formal perceptual study.

Circularity Check

1 steps flagged · score 6.0 of 10

C-19-R SOTA result is forced by test-set hyperparameter selection; other six datasets provide independent support.

  1. fitted input called prediction [Section IV-C2 (beta ablation, Table IX), Section IV-C3 (rank, Table X), Section IV-C4 (patch size, Table XI), and Section IV-B1 (Table II SOTA claim)]
    "In Section IV-B1 the paper claims 'our proposed method achieves the new state-of-the-art (SOTA) compression performance', and Section IV-C2 states: 'To determine its optimal value, we conduct experiments on the C-19-R dataset, evaluating the compression performance for β∈{0.6,0.7,0.8,0.9}. The results demonstrate that β=0.8 provides the optimal trade-off... Therefore, we set β=0.8 as the default value in our method.'"

    The bpp reported for C-19-R in Table II (2.25) is the minimum over the β grid in Table IX, and the same C-19-R test set is used to select the A-LoRA rank (Table X) and patch size (Table XI). Because these hyperparameters were chosen to minimize bpp on C-19-R, the statement that Ours beats DLPR (2.25 vs 2.34) on C-19-R is a restatement of the selection criterion, not an independent prediction. Using a non-optimal β (e.g., 0.7 or 0.9 from Table IX) would give 2.41 or 2.38 bpp, losing to DLPR. The six other datasets were not used for these ablations, so they provide independent support and the circularity is partial.

full rationale

The core derivation is largely self-contained: the adaptive bit-plane decomposition (Eq. 6), dual-path compression, steganography equations (10)-(11), and bitstream accounting (19)-(20) are defined independently of the benchmark outcomes, and the side-information overhead is explicitly included in the bpp. The main circular step is the C-19-R SOTA result: β, rank, and patch size are all selected by minimizing bpp on the C-19-R test set, and the same selected configuration is then reported as SOTA on that dataset, making the 2.25 bpp number an artifact of selection. This is a fitted-input-called-prediction pattern, but it is partial because the six remaining datasets were not used for hyperparameter selection and independently support the method. The citation to [14] (overlapping author K. Chen) for pixel tokenization is ordinary method reuse with external support from [22], not load-bearing circularity. Separately, the paper never reports the message length L used in the Table II compression experiments, so the comparison may charge side information to Ours that baselines do not carry; this is a correctness/fairness risk, not a circularity. Overall score 6.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method relies on standard ML/compression axioms and a set of hyperparameters fitted on the main benchmark. No new physical or conceptual entities are introduced. The slicing index is not a free constant but is derived from a mutual-information threshold controlled by beta, which is itself a fitted hyperparameter.

free parameters (4)
  • beta (information retention hyperparameter) = 0.8
    Controls the split between global and local modalities in Eq. (6); chosen by ablation on C-19-R (Table IX).
  • Patch size P for LLM input = 16
    Determines context length for the LLM; ablated in Table XI, best at 16x16.
  • LoRA rank r and alpha = 64 and 128
    Rank and scaling factor for A-LoRA; ablated in Table X.
  • Slicing index s* = Derived from Eq. (6)
    Determined from mutual information calculated on the training set, but depends on the free parameter beta and the empirical MI estimator.
assumptions (4)
  • standard math Minimizing negative log-likelihood is equivalent to lossless compression (Shannon source coding theorem).
    Used in Section III-A to justify LLM-based compression, Eq. (3).
  • domain assumption Latent variable models cannot capture local fine-grained modalities, while autoregressive models can.
    Assumed in Section III-C to motivate assigning local bit planes to the LLM path; cited from prior works [36], [37].
  • domain assumption Embedding messages in the low-significance bit planes does not alter diagnostic content.
    Used throughout the steganography design (Section III-C2); no clinical validation is provided.
  • domain assumption The empirical mutual information computed over the training set is representative for all test datasets.
    Used in Eq. (6)-(7) to set the slicing index; beta and the MI estimate are global, not per-image.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Joint Lossless Compression and Steganography for Medical Images via Large Language Models." pith.science (2026). https://pith.science/paper/BP3HZPQD

@misc{pith2026250801782,
  author       = {Pith},
  title        = {Pith review of: Joint Lossless Compression and Steganography for Medical Images via Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BP3HZPQD}},
  note         = {Machine review of arXiv:2508.01782}
}
read the original abstract

Recently, large language models (LLMs) have driven promising progress in lossless image compression. However, directly adopting existing paradigms for medical images suffers from an unsatisfactory trade-off between compression performance and efficiency. Moreover, existing LLM-based compressors often overlook the security of the compression process, which is critical in modern medical scenarios. To this end, we propose a novel joint lossless compression and steganography framework. Inspired by bit plane slicing (BPS), we find it feasible to securely embed privacy messages into medical images in an invisible manner. Based on this insight, an adaptive modalities decomposition strategy is first devised to partition the entire image into two segments, providing global and local modalities for subsequent dual-path lossless compression. During this dual-path stage, we innovatively propose a segmented message steganography algorithm within the local modality path to ensure the security of the compression process. Coupled with the proposed anatomical priors-based low-rank adaptation (A-LoRA) fine-tuning strategy, extensive experimental results demonstrate the superiority of our proposed method in terms of compression ratios, efficiency, and security. The source code will be made publicly available.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Exploring Decision-Making Capabilities of LLM Agents: An Experimental Study on Jump-Jump Game

    cs.CV 2025-08 reject novelty 2.0 of 10

    LLM agents can achieve 91% landing success on a simplified Jump-Jump game when prompts include step-by-step reasoning, few-shot examples, and calibration tips, per an unreproducible small study.

Reference graph

Works this paper leans on

75 extracted references · 62 canonical work pages · cited by 1 Pith paper

  1. [1]

    Learned focused plenoptic image compression with microimage preprocessing and global attention,

    K. Tong, X. Jin, Y . Yang, C. Wang, J. Kang, and F. Jiang, “Learned focused plenoptic image compression with microimage preprocessing and global attention,”IEEE Transactions on Multimedia, vol. 26, pp. 890–903, 2023

  2. [2]

    Dnp-aut: Image compression using double-layer non-uniform partition and adaptive u transform,

    Y . Zhang and Z. Cai, “Dnp-aut: Image compression using double-layer non-uniform partition and adaptive u transform,”IEEE Transactions on Multimedia, 2024

  3. [3]

    Learned focused plenoptic image compression with local-global correlation learning,

    G. Liu, H. Yue, B. Wen, and J. Yang, “Learned focused plenoptic image compression with local-global correlation learning,”IEEE Transactions on Multimedia, 2025

  4. [4]

    Multi-stream dense view reconstruction network for light field image compression,

    D. Liu, Y . Huang, Y . Fang, Y . Zuo, and P. An, “Multi-stream dense view reconstruction network for light field image compression,”IEEE Transactions on Multimedia, vol. 25, pp. 4400–4414, 2022

  5. [5]

    Wise: A framework for gigapixel whole-slide-image lossless compression,

    Y . Mao, J. Wang, N. Guan, and C. J. Xue, “Wise: A framework for gigapixel whole-slide-image lossless compression,” inProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025, pp. 29 342–29 351

  6. [6]

    Learned image compression with dictionary-based entropy model,

    J. Lu, L. Zhang, X. Zhou, M. Li, W. Li, and S. Gu, “Learned image compression with dictionary-based entropy model,” inProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025, pp. 12 850–12 859

  7. [7]

    Msgfusion: Medical semantic guided two-branch network for multimodal brain image fusion,

    J. Wen et al., “Msgfusion: Medical semantic guided two-branch network for multimodal brain image fusion,”IEEE Transactions on Multimedia, vol. 26, pp. 944–957, 2023

  8. [8]

    Pathology-preserving transformer based on multi-color space for low-quality medical image enhancement,

    Q. Hou et al., “Pathology-preserving transformer based on multi-color space for low-quality medical image enhancement,”IEEE Transactions on Multimedia, 2025

Show all 75 references
  1. [9]

    Learned lossless im- age compression based on bit plane slicing,

    Z. Zhang, H. Wang, Z. Chen, and S. Liu, “Learned lossless im- age compression based on bit plane slicing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 27 579–27 588

  2. [10]

    Bilateral context modeling for residual coding in lossless 3d medical image compression,

    X. Liu, M. Wang, S. Wang, and S. Kwong, “Bilateral context modeling for residual coding in lossless 3d medical image compression,”IEEE Transactions on Image Processing, vol. 33, pp. 2502–2513, 2024

  3. [11]

    Callic: Content adaptive learning for lossless image compression,

    D. Li, Y . Bai, K. Wang, J. Jiang, X. Liu, and W. Gao, “Callic: Content adaptive learning for lossless image compression,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 2025, pp. 4679– 4688

  4. [12]

    Make lossy compression meaningful for low-light images,

    S. Cai, L. Chen, S. Zhong, L. Yan, J. Zhou, and X. Zou, “Make lossy compression meaningful for low-light images,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 8236– 8245

  5. [13]

    Du et al.,Large language model for lossless image compression with visual prompts, 2025

    J. Du et al.,Large language model for lossless image compression with visual prompts, 2025

  6. [14]

    Chen et al.,Large language models for lossless image compression: Next-pixel prediction in language space is all you need, 2024

    K. Chen et al.,Large language models for lossless image compression: Next-pixel prediction in language space is all you need, 2024

  7. [15]

    D. J. MacKay,Information theory, inference and learning algorithms. Cambridge university press, 2003

  8. [16]

    Com- pression via pre-trained transformers: A study on byte-level multimodal data,

    D. Heurtel-Depeiges, A. Ruoss, J. Veness, and T. Genewein, “Com- pression via pre-trained transformers: A study on byte-level multimodal data,”arXiv preprint arXiv:2410.05078, 2024

  9. [17]

    A low-complexity bit- plane entropy coding and rate control for 3-d dwt based video coding,

    E. Belyaev, K. Egiazarian, and M. Gabbouj, “A low-complexity bit- plane entropy coding and rate control for 3-d dwt based video coding,” IEEE Transactions on Multimedia, vol. 15, no. 8, pp. 1786–1799, 2013

  10. [18]

    Lossless compression of large lan- guage model-generated text via next-token prediction,

    Y . Mao, H. Pirk, and C. J. Xue, “Lossless compression of large lan- guage model-generated text via next-token prediction,”arXiv preprint arXiv:2505.06297, 2025

  11. [19]

    Llmzip: Lossless text compression using large language models,

    C. S. K. Valmeekam, K. Narayanan, D. Kalathil, J.-F. Chamberland, and S. Shakkottai, “Llmzip: Lossless text compression using large language models,”arXiv preprint arXiv:2306.04050, 2023

  12. [20]

    Multimedia selective encryp- tion by means of randomized arithmetic coding,

    M. Grangetto, E. Magli, and G. Olmo, “Multimedia selective encryp- tion by means of randomized arithmetic coding,”IEEE Transactions on Multimedia, vol. 8, no. 5, pp. 905–917, 2006

  13. [21]

    Cryptanalysis of some multime- dia encryption schemes,

    G. Jakimoski and K. Subbalakshmi, “Cryptanalysis of some multime- dia encryption schemes,”IEEE transactions on multimedia, vol. 10, no. 3, pp. 330–338, 2008

  14. [22]

    Language modeling is compression,

    G. Del ´etang et al., “Language modeling is compression,” inICLR, 2024

  15. [23]

    Variational autoencoder,

    L. Pinheiro Cinelli, M. Ara ´ujo Marins, E. A. Barros da Silva, and S. Lima Netto, “Variational autoencoder,” inVariational methods for machine learning with applications to deep networks, Springer, 2021, pp. 111–149

  16. [24]

    Child,Very deep vaes generalize autoregressive models and can outperform them on images, 2021

    R. Child,Very deep vaes generalize autoregressive models and can outperform them on images, 2021

  17. [25]

    Initial bits generation for lossless image compression using bits-back coding with a latent variable model,

    H. Sugiyama, K. Suzuki, and I. Matsuda, “Initial bits generation for lossless image compression using bits-back coding with a latent variable model,” inInternational Workshop on Advanced Imaging Technology (IWAIT) 2024, SPIE, vol. 13164, 2024, pp. 551–555

  18. [26]

    Variational autoencoder for low bit-rate image compression,

    L. Zhou, C. Cai, Y . Gao, S. Su, and J. Wu, “Variational autoencoder for low bit-rate image compression,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018

  19. [27]

    Generative latent coding for ultra-low bitrate image compression,

    Z. Jia, J. Li, B. Li, H. Li, and Y . Lu, “Generative latent coding for ultra-low bitrate image compression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 26 088–26 098

  20. [28]

    Visual autoregres- sive modeling: Scalable image generation via next-scale prediction,

    K. Tian, Y . Jiang, Z. Yuan, B. Peng, and L. Wang, “Visual autoregres- sive modeling: Scalable image generation via next-scale prediction,” arXiv preprint arXiv:2404.02905, 2024

  21. [29]

    Sentencevae: Enable next-sentence prediction for large language models with faster speed, higher accuracy and longer context,

    H. An, Y . Chen, Z. Sun, and X. Li, “Sentencevae: Enable next-sentence prediction for large language models with faster speed, higher accuracy and longer context,”arXiv preprint arXiv:2408.00655, 2024

  22. [30]

    Multi- direction dictionary learning based depth map super-resolution with JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14 autoregressive modeling,

    J. Wang, W. Xu, J.-F. Cai, Q. Zhu, Y . Shi, and B. Yin, “Multi- direction dictionary learning based depth map super-resolution with JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14 autoregressive modeling,”IEEE Transactions on Multimedia, vol. 22, no. 6, pp. 1470–1...

  23. [31]

    A unified traffic model for mpeg- 4 and h. 264 video traces,

    M. Dai, Y . Zhang, and D. Loguinov, “A unified traffic model for mpeg- 4 and h. 264 video traces,”IEEE Transactions on Multimedia, vol. 11, no. 5, pp. 1010–1023, 2009

  24. [32]

    Pilc: Practical image lossless compression with an end-to-end gpu oriented neural framework,

    N. Kang, S. Qiu, S. Zhang, Z. Li, and S.-T. Xia, “Pilc: Practical image lossless compression with an end-to-end gpu oriented neural framework,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 3739–3748

  25. [33]

    Managing dicom images: Tips and tricks for the radiologist,

    D. R. Varma, “Managing dicom images: Tips and tricks for the radiologist,”Indian Journal of Radiology and Imaging, vol. 22, no. 01, pp. 4–13, 2012

  26. [34]

    Auxiliary metadata delivery in view synthesis using depth no synthesis error model,

    S.-C. Pei and Y .-Y . Wang, “Auxiliary metadata delivery in view synthesis using depth no synthesis error model,”IEEE Transactions on Multimedia, vol. 17, no. 1, pp. 128–133, 2014

  27. [35]

    Multiple-image encryption with bit-plane decomposition and chaotic maps,

    Z. Tang, J. Song, X. Zhang, and R. Sun, “Multiple-image encryption with bit-plane decomposition and chaotic maps,”Optics and Lasers in Engineering, vol. 80, pp. 1–11, 2016

  28. [36]

    Understanding anomaly detection with deep invertible networks through hierarchies of distributions and features,

    R. Schirrmeister, Y . Zhou, T. Ball, and D. Zhang, “Understanding anomaly detection with deep invertible networks through hierarchies of distributions and features,”Advances in Neural Information Processing Systems, vol. 33, pp. 21 038–21 049, 2020

  29. [37]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,”The Bell system technical journal, vol. 27, no. 3, pp. 379–423, 1948

  30. [38]

    Practical full resolution learned lossless image compression,

    F. Mentzer, E. Agustsson, M. Tschannen, R. Timofte, and L. V . Gool, “Practical full resolution learned lossless image compression,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 10 629–10 638

  31. [39]

    Learning scalableℓ ∞- constrained near-lossless image compression via joint lossy image and residual compression,

    Y . Bai, X. Liu, W. Zuo, Y . Wang, and X. Ji, “Learning scalableℓ ∞- constrained near-lossless image compression via joint lossy image and residual compression,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

  32. [40]

    How well can vision language models see image details?

    C. Gou et al., “How well can vision language models see image details?”arXiv preprint arXiv:2408.03940, 2024

  33. [41]

    An improved baseline for reasoning segmentation with large language model,

    S. Yang et al., “An improved baseline for reasoning segmentation with large language model,”CoRR, 2023

  34. [42]

    Compression represents intelligence linearly,

    Y . Huang, J. Zhang, Z. Shan, and J. He, “Compression represents intelligence linearly,”arXiv preprint arXiv:2404.09937, 2024

  35. [43]

    Png (portable network graphics) specification version 1.0,

    T. Boutell, “Png (portable network graphics) specification version 1.0,” W3C, Tech. Rep., 1997

  36. [44]

    Hiding data in images by simple lsb substitution,

    C.-K. Chan and L.-M. Cheng, “Hiding data in images by simple lsb substitution,”Pattern recognition, vol. 37, no. 3, pp. 469–474, 2004

  37. [45]

    Designing steganographic distortion using directional filters,

    V . Holub and J. Fridrich, “Designing steganographic distortion using directional filters,” in2012 IEEE International workshop on informa- tion forensics and security (WIFS), IEEE, 2012, pp. 234–239

  38. [46]

    Generative steganography via auto-generation of semantic object contours,

    Z. Zhou et al., “Generative steganography via auto-generation of semantic object contours,”IEEE Transactions on Information Forensics and Security, vol. 18, pp. 2751–2765, 2023

  39. [47]

    Secret-to-image reversible transformation for gener- ative steganography,

    Z. Zhou et al., “Secret-to-image reversible transformation for gener- ative steganography,”IEEE Transactions on Dependable and Secure Computing, vol. 20, no. 5, pp. 4118–4134, 2022

  40. [48]

    Generative steganography based on long readable text generation,

    Y . Cao et al., “Generative steganography based on long readable text generation,”IEEE transactions on computational social systems, vol. 11, no. 4, pp. 4584–4594, 2022

  41. [49]

    Establishing robust generative image steganography via popular stable diffusion,

    X. Hu, S. Li, Q. Ying, W. Peng, X. Zhang, and Z. Qian, “Establishing robust generative image steganography via popular stable diffusion,” IEEE Transactions on Information Forensics and Security, vol. 19, pp. 8094–8108, 2024

  42. [50]

    Stegafds: Generative steganography based on first-order dpm-solver,

    C. Li, W. Li, Z. Xu, and N. Yu, “Stegafds: Generative steganography based on first-order dpm-solver,” in2024 IEEE 23rd International Conference on Trust, Security and Privacy in Computing and Com- munications (TrustCom), IEEE, 2024, pp. 109–116

  43. [51]

    Coding theorems for shannon’s cipher system with cor- related source outputs, and common information,

    H. Yamamoto, “Coding theorems for shannon’s cipher system with cor- related source outputs, and common information,”IEEE Transactions on Information Theory, vol. 40, no. 1, pp. 85–95, 1994

  44. [52]

    Summary of the hipaa privacy rule,

    H. C. Assistance, “Summary of the hipaa privacy rule,”Office for Civil Rights, 2003

  45. [53]

    Implementation of lsb steganog- raphy and its evaluation for various bits,

    D. Neeta, K. Snehal, and D. Jacobs, “Implementation of lsb steganog- raphy and its evaluation for various bits,” in2006 1st international conference on digital information management, IEEE, 2006, pp. 173– 178

  46. [54]

    The transport layer security (tls) protocol version 1.3,

    E. Rescorla, “The transport layer security (tls) protocol version 1.3,” Tech. Rep., 2018

  47. [55]

    Security architecture for the internet protocol,

    S. Kent and K. Seo, “Security architecture for the internet protocol,” Tech. Rep., 2005

  48. [56]

    Using high-dimensional image models to perform highly undetectable steganography,

    T. Pevn `y, T. Filler, and P. Bas, “Using high-dimensional image models to perform highly undetectable steganography,” inInternational workshop on information hiding, Springer, 2010, pp. 161–177

  49. [57]

    Universal distortion function for steganography in an arbitrary domain,

    V . Holub, J. Fridrich, and T. Denemark, “Universal distortion function for steganography in an arbitrary domain,”EURASIP Journal on Information Security, vol. 2014, no. 1, p. 1, 2014

  50. [58]

    Rich models for steganalysis of digital images,

    J. Fridrich and J. Kodovsky, “Rich models for steganalysis of digital images,”IEEE Transactions on information Forensics and Security, vol. 7, no. 3, pp. 868–882, 2012

  51. [59]

    Lora: Low-rank adaptation of large language models.,

    E. J. Hu et al., “Lora: Low-rank adaptation of large language models.,” ICLR, vol. 1, no. 2, p. 3, 2022

  52. [60]

    Searching for mobilenetv3,

    A. Howard et al., “Searching for mobilenetv3,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1314–1324

  53. [61]

    Lossless compression of raw images by learning the prediction and frequency decomposition,

    H. Rhee and N. I. Cho, “Lossless compression of raw images by learning the prediction and frequency decomposition,” in2023 IEEE International Conference on Visual Communications and Image Pro- cessing (VCIP), 2023, pp. 1–5

  54. [62]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

  55. [63]

    Sgdr: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “Sgdr: Stochastic gradient descent with warm restarts,”arXiv preprint arXiv:1608.03983, 2016

  56. [64]

    The loco-i lossless image compression algorithm: Principles and standardization into jpeg- ls,

    M. J. Weinberger, G. Seroussi, and G. Sapiro, “The loco-i lossless image compression algorithm: Principles and standardization into jpeg- ls,”IEEE Transactions on Image processing, vol. 9, no. 8, pp. 1309– 1324, 2000

  57. [65]

    An overview of jpeg-2000,

    M. W. Marcellin, M. J. Gormish, A. Bilgin, and M. P. Boliek, “An overview of jpeg-2000,” inProceedings DCC 2000. Data compression conference, IEEE, 2000, pp. 523–541

  58. [66]

    Research on the webp image format,

    Z. Si and K. Shen, “Research on the webp image format,” inAd- vanced graphic communications, packaging technology and materials, Springer, 2015, pp. 271–277

  59. [67]

    Flif: Free lossless image format based on maniac compression,

    J. Sneyers and P. Wuille, “Flif: Free lossless image format based on maniac compression,” in2016 IEEE international conference on image processing (ICIP), IEEE, 2016, pp. 66–70

  60. [68]

    High bit-depth medical image compression with hevc,

    S. S. Parikh, D. Ruiz, H. Kalva, G. Fern ´andez-Escribano, and V . Adzic, “High bit-depth medical image compression with hevc,”IEEE journal of biomedical and health informatics, vol. 22, no. 2, pp. 552–560, 2017

  61. [69]

    Jpeg xl next-generation image compression archi- tecture and coding tools,

    J. Alakuijala et al., “Jpeg xl next-generation image compression archi- tecture and coding tools,” inApplications of digital image processing XLII, SPIE, vol. 11137, 2019, pp. 112–124

  62. [70]

    Overview of the versatile video coding (vvc) standard and its applications,

    B. Bross et al., “Overview of the versatile video coding (vvc) standard and its applications,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 10, pp. 3736–3764, 2021

  63. [71]

    Lc-fdnet: Learned lossless image compression with frequency decomposition network,

    H. Rhee, Y . I. Jang, S. Kim, and N. I. Cho, “Lc-fdnet: Learned lossless image compression with frequency decomposition network,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 6023–6032

  64. [72]

    Learning lossless compression for high bit-depth volumetric medical image,

    K. Wang, Y . Bai, D. Li, D. Zhai, J. Jiang, and X. Liu, “Learning lossless compression for high bit-depth volumetric medical image,” IEEE Transactions on Image Processing, 2024

  65. [73]

    Deep lossy plus residual coding for lossless and near-lossless image compression,

    Y . Bai, X. Liu, K. Wang, X. Ji, X. Wu, and W. Gao, “Deep lossy plus residual coding for lossless and near-lossless image compression,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 5, pp. 3577–3594, 2024

  66. [74]

    Secure, privacy-preserving and federated machine learning in medical imaging,

    G. A. Kaissis, M. R. Makowski, D. R ¨uckert, and R. F. Braren, “Secure, privacy-preserving and federated machine learning in medical imaging,”Nature Machine Intelligence, vol. 2, no. 6, pp. 305–311, 2020

  67. [75]

    Deep residual network for steganalysis of digital images,

    M. Boroumand, M. Chen, and J. Fridrich, “Deep residual network for steganalysis of digital images,”IEEE Transactions on Information Forensics and Security, vol. 14, no. 5, pp. 1181–1193, 2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.