pith. sign in

arxiv: 1907.04362 · v1 · pith:ZJWYC3CMnew · submitted 2019-07-09 · 💻 cs.CV · cs.MM

BASN -- Learning Steganography with Binary Attention Mechanism

Yang Yang This is my paper

Pith reviewed 2026-05-25 00:15 UTC · model grok-4.3

classification 💻 cs.CV cs.MM
keywords image steganographybinary attention mechanismsteganalysis resistancefeature map preservationpayload capacityconvolutional neural networkssecret data embedding
0
0 comments X

The pith

A binary attention mechanism embeds secret data in images while keeping task feature maps unrelated to the hidden information.

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

The paper tries to establish that adding a binary attention mechanism to image steganography models lets them carry more secret data without making the feature maps of task-specific networks responsive to that data. A sympathetic reader would care because convolutional networks now dominate image analysis on the internet, so any method that hides information without triggering those networks offers a practical way to share secrets more securely. If the claim holds, steganography could support higher payloads in everyday image carriers while still passing undetected by current automated steganalysis tools and without degrading performance on classification or detection tasks.

Core claim

Introducing a binary attention mechanism into steganography allows the model to embed secret information such that task-specific feature maps remain irrelative to the hidden data, which produces high payload capacity, little feature map distortion, and resistance to state-of-the-art image steganalysis algorithms.

What carries the argument

Binary attention mechanism, which directs embedding to preserve irrelevance between hidden data and task-specific feature maps.

If this is right

  • Higher payload capacity becomes possible in image carriers without increasing feature map distortion.
  • Task-specific networks continue to produce the same outputs on stego images as on cover images.
  • The embedded data remains undetected by existing steganalysis algorithms that rely on neural networks.
  • Secret sharing through internet images can occur at larger scales while automated vision pipelines remain unaffected.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same attention control on feature relevance might reduce the usual capacity-security trade-off in other embedding tasks.
  • The approach could be tested on video or audio carriers to check whether the irrelevance property transfers across media types.
  • If the mechanism works across many task networks, it points toward attention layers as a general design tool for making embeddings invisible to downstream models.

Load-bearing premise

Keeping task-specific feature maps unrelated to the embedded secret data will directly improve resistance to neural-network steganalysis.

What would settle it

Train a fresh steganalysis network on a large set of cover and stego images produced by the binary-attention method and measure whether its detection accuracy exceeds what random guessing would achieve.

Figures

Figures reproduced from arXiv: 1907.04362 by Yang Yang.

Figure 1
Figure 1. Figure 1: LSB-Matching Embedded Image Misclassification [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The Embedding and Extraction Architecture [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The gradient impact comparison between variance and variance pooling [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Image Smoothing Effect Comparison minimize Vitc(Aitc · Cθ + (1 − Aitc) · C) (5) subject to 1 N X N i Aitc ≤ θ (6) The θ in Equation 6 is used as an upper bound to limit down the attention area size. If trained without it, model fitc is free to output all-ones matrix Aitc to acquire an optimal texture-free image. It is well-known that an image with the least amount of texture is a solid color image, which d… view at source ↗
Figure 5
Figure 5. Figure 5: The Effect of ITC Attention on Texture Complexity Reduction [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Model Architectures Here, C stands for the cover image and S stands for the corresponding embedded image. Lfmrl(·) is the feature map reconstruction loss and α, β are thresholds limiting the area of attention map acting the same role as θ in the ITC attention model. The actual ways of training the MFD attention model is split into 2 phases (See [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: MFD Attention Mechanism Training Pipeline [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The Encoder and Decoder Block of the MFD Attention Model [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Soft Area Penalties (a) The Cover (b) MFD Attention (c) The Embedded (d) The Cover (e) MFD Attention (f) The Embedded [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The Visual Effect of MFD Attention on Embedding with Random Noise [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The 1st Phase Finetune Pipeline 3 Fusion Strategies, Finetune Process and In￾ference Techniques The fusion strategies help merge ITC and MFD attention models into one attention model, and thus they are substantial to be consistent and stable. In this paper, two fusion strategies being minima fusion and mean fusion are put forth as Equation 15 and 16. Minima fusion strategy aims to improve security while m… view at source ↗
Figure 12
Figure 12. Figure 12: ITC Attention After Finetune The first column shows the original image, the second column shows the ITC attention before any finetune, the third column shows the ITC attention after finetuning for minima fusion strategy, and the forth column shows the ITC attention after finetuning for mean fusion strategy. named Least Significant Masking (LSM) which masks the lowest several bits of the attention during e… view at source ↗
Figure 13
Figure 13. Figure 13: Steganography using Mean Fusion with 1-bit LSM [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: ROC Curves: Steganalysis with StegExpose, SPAM Features, SRM [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: ResNet-18 Classification Feature Distortion Rate [PITH_FULL_IMAGE:figures/full_fig_p016_15.png] view at source ↗
read the original abstract

Secret information sharing through image carrier has aroused much research attention in recent years with images' growing domination on the Internet and mobile applications. However, with the booming trend of convolutional neural networks, image steganography is facing a more significant challenge from neural-network-automated tasks. To improve the security of image steganography and minimize task result distortion, models must maintain the feature maps generated by task-specific networks being irrelative to any hidden information embedded in the carrier. This paper introduces a binary attention mechanism into image steganography to help alleviate the security issue, and in the meanwhile, increase embedding payload capacity. The experimental results show that our method has the advantage of high payload capacity with little feature map distortion and still resist detection by state-of-the-art image steganalysis algorithms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes BASN, a binary attention mechanism for image steganography. It aims to embed secret data in carrier images such that task-specific feature maps remain irrelative to the hidden payload, thereby achieving high embedding capacity, minimal feature-map distortion, and resistance to state-of-the-art neural steganalysis detectors.

Significance. If validated, the approach of using binary attention to enforce feature-map irrelativity could offer a principled way to improve robustness against CNN-based steganalyzers while preserving payload. The emphasis on decoupling task features from embedding artifacts addresses a timely concern in adversarial steganography.

major comments (2)
  1. [Abstract] Abstract: the central claim that 'the experimental results show that our method has the advantage of high payload capacity with little feature map distortion and still resist detection by state-of-the-art image steganalysis algorithms' supplies no numerical metrics, baselines, datasets, or specific detector names, rendering the empirical support for the core security and capacity assertions unverifiable.
  2. [Abstract] The manuscript's key assumption—that binary attention will keep task-specific feature maps irrelative to embedded data and that this irrelativity will causally produce resistance to NN steganalysis—is presented without supporting measurements (e.g., feature-map correlation coefficients, ablation on attention masks, or similarity metrics) or analysis of alternative statistical artifacts that steganalyzers might exploit.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments on the abstract below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'the experimental results show that our method has the advantage of high payload capacity with little feature map distortion and still resist detection by state-of-the-art image steganalysis algorithms' supplies no numerical metrics, baselines, datasets, or specific detector names, rendering the empirical support for the core security and capacity assertions unverifiable.

    Authors: We agree that the abstract would be strengthened by including concrete details. In the revised manuscript we will expand the abstract to report specific payload capacities (in bpp), quantitative distortion measures on feature maps, the datasets employed, and the names of the steganalysis detectors against which resistance was evaluated. revision: yes

  2. Referee: [Abstract] The manuscript's key assumption—that binary attention will keep task-specific feature maps irrelative to embedded data and that this irrelativity will causally produce resistance to NN steganalysis—is presented without supporting measurements (e.g., feature-map correlation coefficients, ablation on attention masks, or similarity metrics) or analysis of alternative statistical artifacts that steganalyzers might exploit.

    Authors: The binary attention mechanism is explicitly constructed to enforce feature-map irrelativity; the manuscript demonstrates the downstream effect through end-to-end steganalysis resistance results. We will revise the abstract to point more explicitly to these empirical outcomes. Adding new quantitative analyses such as correlation coefficients or mask ablations would require supplementary experiments beyond the current submission. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical ML proposal without derivation chain

full rationale

The paper proposes an empirical architecture (binary attention for steganography) whose central claims rest on experimental outcomes against steganalysis detectors rather than any mathematical derivation, prediction, or first-principles result. No equations, fitted parameters renamed as predictions, or self-citation load-bearing steps appear in the abstract or described content. The method is therefore self-contained against external benchmarks and receives the default non-finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities.

pith-pipeline@v0.9.0 · 5647 in / 912 out tokens · 17335 ms · 2026-05-25T00:15:26.874387+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 2 internal anchors

  1. [1]

    Hiding images in plain sight: Deep steganography

    Shumeet Baluja. Hiding images in plain sight: Deep steganography. In Advances in Neural Information Processing Systems , pages 2069–2079, 2017

    work page 2069

  2. [2]

    StegExpose - A Tool for Detecting LSB Steganography

    Benedikt Boehm. StegExpose - A Tool for Detecting LSB Steganogra- phy. arXiv e-prints, 2014. arXiv: 1410.6656

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  3. [3]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. 2009. 16

    work page 2009

  4. [4]

    Rich models for steganalysis of dig- ital images

    Jessica Fridrich and Jan Kodovsky. Rich models for steganalysis of dig- ital images. IEEE Transactions on Information Forensics and Security , 7(3):868–882, 2012

    work page 2012

  5. [5]

    Fast r-cnn

    Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440–1448, 2015

    work page 2015

  6. [6]

    The complete works of william shakespeare by william shakespeare - free ebook., 2018

    Project Gutenberg. The complete works of william shakespeare by william shakespeare - free ebook., 2018. [Online; Accessed 13-Nov-2018]

    work page 2018

  7. [7]

    Project gutenberg, 2018

    Project Gutenberg. Project gutenberg, 2018. [Online; Accessed 13-Nov- 2018]

    work page 2018

  8. [8]

    Deep resid- ual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep resid- ual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition , pages 770–778, 2016

    work page 2016

  9. [9]

    Universal dis- tortion function for steganography in an arbitrary domain

    Vojtˇ ech Holub, Jessica Fridrich, and Tom´ aˇ s Denemark. Universal dis- tortion function for steganography in an arbitrary domain. EURASIP Journal on Information Security , 2014(1):1, 2014

    work page 2014

  10. [10]

    A novel image steganography method via deep convolutional generative adversarial networks

    Donghui Hu, Liang Wang, Wenjie Jiang, Shuli Zheng, and Bin Li. A novel image steganography method via deep convolutional generative adversarial networks. IEEE Access, 6:38303–38314, 2018

    work page 2018

  11. [11]

    Adam: A Method for Stochastic Optimization

    Diederik Kingma and Jimmy Ba. Adam: A Method for Stochastic Op- timization. arXiv e-prints, 2014. arXiv:1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  12. [12]

    Mielikainen

    J. Mielikainen. Lsb matching revisited. IEEE signal processing letters , 13(5):285–287, 2006

    work page 2006

  13. [13]

    Steganalysis by subtrac- tive pixel adjacency matrix

    Tom´ aˇ s Pevny, Patrick Bas, and Jessica Fridrich. Steganalysis by subtrac- tive pixel adjacency matrix. IEEE Transactions on information Foren- sics and Security , 5(2):215–224, 2010

    work page 2010

  14. [14]

    Using high-dimensional image models to perform highly undetectable steganography

    Tom´ aˇ s Pevn` y, Tom´ aˇ s Filler, and Patrick Bas. Using high-dimensional image models to perform highly undetectable steganography. In In- ternational Workshop on Information Hiding , pages 161–177. Springer, 2010

    work page 2010

  15. [15]

    You only look once: Unified, real-time object detection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 779–788, 2016. 17

    work page 2016

  16. [16]

    Facenet: A unified embedding for face recognition and clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 815–823, 2015

    work page 2015

  17. [17]

    Shakespeare

    W. Shakespeare. The Complete Works of William Shakespeare . 1994

    work page 1994

  18. [18]

    Bpcs steganography using ezw lossy compressed images

    Jeremiah Spaulding, Hideki Noda, Mahdad N Shirazi, and Eiji Kawaguchi. Bpcs steganography using ezw lossy compressed images. Pattern Recognition Letters, 23(13):1579–1587, 2002

    work page 2002

  19. [19]

    A new information hiding method based on improved bpcs steganography

    Shuliang Sun. A new information hiding method based on improved bpcs steganography. Advances in Multimedia, 2015:5, 2015

    work page 2015

  20. [20]

    On the importance of initialization and momentum in deep learning

    Ilya Sutskever, James Martens, George Dahl, and Geoffrey Hinton. On the importance of initialization and momentum in deep learning. In International conference on machine learning , pages 1139–1147, 2013

    work page 2013

  21. [21]

    Inception-v4, inception-resnet and the impact of residual con- nections on learning

    Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, inception-resnet and the impact of residual con- nections on learning. In Thirty-First AAAI Conference on Artificial Intelligence, 2017

    work page 2017

  22. [22]

    A high quality steganographic method with pixel-value dif- ferencing and modulus function

    Chung-Ming Wang, Nan-I Wu, Chwei-Shyong Tsai, and Min-Shiang Hwang. A high quality steganographic method with pixel-value dif- ferencing and modulus function. Journal of Systems and Software , 81(1):150–158, 2008

    work page 2008

  23. [23]

    F5—a steganographic algorithm

    Andreas Westfeld. F5—a steganographic algorithm. In Ira S. Moskowitz, editor, Information Hiding , pages 289–302, Berlin, Heidelberg, 2001. Springer Berlin Heidelberg

    work page 2001

  24. [24]

    A steganographic method for images by pixel-value differencing

    Da-Chun Wu and Wen-Hsiang Tsai. A steganographic method for images by pixel-value differencing. Pattern Recognition Letters, 24(9- 10):1613–1626, 2003

    work page 2003

  25. [25]

    Image steganographic scheme based on pixel-value differencing and lsb replacement methods

    H-C Wu, N-I Wu, C-S Tsai, and M-S Hwang. Image steganographic scheme based on pixel-value differencing and lsb replacement methods. IEE Proceedings-Vision, Image and Signal Processing , 152(5):611–615, 2005

    work page 2005

  26. [26]

    Image-into-image steganogra- phy using deep convolutional network

    Pin Wu, Yang Yang, and Xiaoqiang Li. Image-into-image steganogra- phy using deep convolutional network. In Richang Hong, Wen-Huang Cheng, Toshihiko Yamasaki, Meng Wang, and Chong-Wah Ngo, editors, 18 Advances in Multimedia Information Processing – PCM 2018 , pages 792–802, Cham, 2018. Springer International Publishing

    work page 2018

  27. [27]

    Yedroudj-net: An efficient cnn for spatial steganalysis

    Mehdi Yedroudj, Fr´ ed´ eric Comby, and Marc Chaumont. Yedroudj-net: An efficient cnn for spatial steganalysis. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2092–2096. IEEE, 2018

    work page 2018

  28. [28]

    Just-noticeable difference estimation with pixels in images

    Xiaohui Zhang, Weisi Lin, and Ping Xue. Just-noticeable difference estimation with pixels in images. Journal of Visual Communication and Image Representation, 19(1):30–41, 1 2008

    work page 2008

  29. [29]

    Camera style adaptation for person re-identification

    Zhun Zhong, Liang Zheng, Zhedong Zheng, Shaozi Li, and Yi Yang. Camera style adaptation for person re-identification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 5157–5166, 2018. 19

    work page 2018