Learning-Based Interface for Semantic Communication with Bit Importance Awareness

Wenyi Zhang; Wenzheng Kong

REVIEW 3 major objections 10 minor 17 references

Learning-Based Interface for Semantic Communication with Bit Importance Awareness

T0 review · 3 major / 10 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Learned bit-importance interface lifts wireless image PSNR

desk verdict Incremental but promising extension of Split DeepJSCC, undermined by an unfair baseline comparison: the proposed model is trained over 5–20 dB while Split DeepJSCC is trained only at 20 dB. read the letter →

arxiv 2507.12850 v1 pith:JBRPBWZX submitted 2025-07-17 cs.IT cs.NImath.IT

Wenzheng Kong , Wenyi Zhang This is my paper · ORCID

classification cs.ITcs.NImath.IT

keywords semanticcommunicationjointsource-channelcodinglearnablebinaryinterfacebitimportancewirelessimagetransmissiondeeplearningtwo-stagetrainingsplitJSCC

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

Joint source-channel coding typically trains source and channel as one end-to-end system, which is hard to deploy in today's networks where different providers run the two ends. This paper claims that the bit-level interface separating them can be made learnable rather than prescribed: the interface's bit-flipping probabilities are trained alongside the source code, and each bit's learned importance is passed to the channel code. An Importance-Aware Net uses that importance signal to focus channel coding on bits that matter for image quality. In tests on wireless image transmission, the approach improves PSNR over the fixed-interface Split DeepJSCC baseline across datasets, bandwidth ratios, and both AWGN and Rayleigh channels, with the biggest gains at high channel bandwidth ratios. A working scheme of this kind would let semantic communication fit into existing layered wireless networks.

What carries the argument

The load-bearing object is the trainable binary interface: an array of M binary symmetric channels with parameters {ϵ_n}, each describing the flip probability of one bit produced by the source mapper. The source stage maximizes variational mutual information between image s and noisy bits b̂ through the noise distribution q_noisy, with straight-through estimation for gradients; the regularization term λΣ(ϵ_n−0.5)² prevents collapse to error-free bits. The learned ϵ values then double as bit-importance scores, mapped to attention weights through 1−2ϵ in the Importance-Aware Net, which combines interface attention with a squeeze-and-excitation block inside the channel mapper and demapper. This chain links the source code's quality view to the channel code's resource allocation without joint training.

What would settle it

Inspect the learned bit-flip probabilities after stage one and compare them against an empirical sensitivity map: corrupt each bit of the frozen source encoder's output over the real AWGN and Rayleigh channels and measure the PSNR drop. If the bits the interface ranks most important are not the bits whose corruption hurts reconstruction most, the transfer premise fails and the Importance-Aware Net would be attending to the wrong positions.

Watch

Extended reading notes

Core claim

The core claim is that the binary interface—the shared bit-level agreement between source and network sides—should be learned, not fixed. The paper trains an array of binary symmetric channels with per-bit flip probabilities ϵ_n during the source stage, so the source mapper learns which bit positions the noisy pipe will respect; a regularizer keeps the flipping probabilities away from zero. After freezing the source code, the fixed ϵ values are transformed to 1−2ϵ and injected into the channel mapper and demapper through an Importance-Aware Net, telling the channel code which bits are most valuable for reconstruction. The reported result is that this two-stage split training with learned importance outperforms Split DeepJSCC in PSNR over AWGN and Rayleigh channels across four datasets and all tested bandwidth ratios and signal-to-noise ratios.

Load-bearing premise

The learned ranking of bit importance, produced while training the source code through simple binary symmetric channels, must remain meaningful when the actual transmission path is an AWGN or Rayleigh wireless channel, since that ordering is what steers the channel code.

Editorial extensions

If this is right

Source and channel codes can be trained separately and frozen at different times, while the learned interface keeps them coordinated—a property needed when source and network are operated by different entities.
Performance gains grow with channel bandwidth ratio, so the scheme exploits available bandwidth better than a fixed interface that saturates as the bandwidth ratio rises.
Under Rayleigh fading, the method degrades more slowly at low SNR and gains more at high SNR, making it suited to time-varying wireless conditions.
The model trained only on CIFAR-10 still improves PSNR on CIFAR-100, suggesting the learned interface is not overtuned to a single dataset.
Each deployed pair of source and channel codes still fits the layered network architecture, so no joint end-to-end retraining is required.

Reading between the lines

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

The learned ϵ values could be exported as an explicit unequal-error-protection profile, letting a standard channel code allocate stronger protection to the top-ranked bits without retraining the source side.
If the bit-importance ordering is stable across images, the same interface could be trained once on generic data and reused for unseen sources, making the split design closer to a drop-in network feature.
The importance signal might extend to other modalities, such as text or video, where the same two-stage interface could define which bits carry semantic weight.
A testable next step is to vary the regularizer strength and measure the trade-off: weaker regularization should sharpen the importance ranking but may make the source bits less error-resilient.

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.

Desk Editor's Note

read the letter

The paper does something genuinely useful: it treats the BSC bit-flipping probabilities at the interface as trainable, then feeds the learned importance (1−2ε) into an attention module for the channel mapper/demapper. That combination is new as far as I can tell, and the two-stage training with frozen source codec is a clean way to separate source and channel design. The math is standard but solid: variational MI, STE, and the BSC model are all handled correctly. The ablation study, which shows that removing the IAN or just the interface attention degrades PSNR, gives real evidence that the importance-aware mechanism itself helps.

The soft spot is exactly what the stress-test note flags, and it is not minor. Section III-A.4 states that the proposed model is trained with SNR uniformly sampled from 5 to 20 dB, while the Split DeepJSCC baseline is trained at a fixed 20 dB SNR. The evaluation then reports PSNR across 5–20 dB. A baseline trained only at 20 dB will naturally underperform at 10 dB or 5 dB, even if the architectures were identical. So the headline claim—"outperforms Split DeepJSCC across all CBRs and SNRs"—is not established by the figures. The ablation does not fix this, because it compares the proposed variants (all trained on 5–20 dB) against the same fixed-20 dB baseline.

There are two other issues that need to be addressed, though they are less central. First, no error bars or statistical significance are reported; the curves look clean, but we have no sense of variance. Second, the paper claims "dynamical adaptation to diverse channel bandwidth ratios," but it never explains how a model trained for one CBR handles another. Are separate models trained per CBR, or does the IAN take CBR as an input? The text is vague, and the reader is left guessing.

On the positive side, the paper is honest about its debts: it explicitly cites [9] for trainable bit-flipping probabilities and [13] for the ChannelModNet architecture. The code is promised, which is good. The central idea—letting the interface learn which bits matter and using that to guide channel coding—is worth exploring, and the ablation suggests the IAN has real value.

Who should read this? People working on split JSCC or semantic communication interfaces will find the design worth considering, but they should be careful about accepting the headline results without a matched-training baseline. This paper deserves a serious referee, but the referee should insist on a rerun with the baseline trained on the same SNR distribution, plus error bars and a clear statement of how CBR adaptation works.

Referee Report

3 major / 10 minor

Summary. The paper proposes a learning-based binary interface for the Split DeepJSCC framework in wireless image transmission. The interface is parameterized by bit-flipping probabilities {epsilon_n} of an array of BSCs, which are trained jointly with the source mapper and demapper in a first stage via a variational mutual-information objective with a regularization term. In a second stage, the trained epsilon values are used as bit-importance features in an Importance-Aware Net (IAN) that conditions the channel mapper and demapper on bit importance. The authors report PSNR improvements over Split DeepJSCC across four datasets, four channel bandwidth ratios, and AWGN/Rayleigh channels. The claimed contribution is a trainable interface that eliminates prescribed reliability configurations and provides useful importance information to the channel code.

Significance. If the reported gains are robust, the paper is a useful incremental step toward making DeepJSCC compatible with existing layered network architectures. The formulation is clean: the variational MI objective with BSC noise and Straight-Through Estimator is standard, the regularization is sensible, and the IAN is a reasonable way to inject bit-importance information. The code is made publicly available, which supports reproducibility. The main significance currently rests on the experimental comparison to Split DeepJSCC, which is compromised by an asymmetric training protocol (see major comments). The internal ablation study is better controlled and provides some evidence that the IAN contributes positively, but it covers only one dataset and one CBR.

major comments (3)

[Section III-A.4 and Figs. 5-6] The ablation study in Section III-C does not remove this confound, because it does not include a matched-training Split DeepJSCC baseline; all four schemes in the ablation use the proposed training SNR distribution, so the comparison to Split DeepJSCC in Fig. 7 inherits the same asymmetry.
[Figs. 5-6 (Section III-B)] At SNR = 20 dB, the comparison is fair because the baseline is trained at its operating point; the confound primarily affects the lower and middle SNR range. A matched-SNR-distribution baseline would clarify the magnitude of the true improvement.
[Section II-A and Section II-C] This issue does not invalidate the empirical results, but it affects the practical decoupling motivation that is central to the paper's framing.

minor comments (10)

[Section II-A] Typo: 'wirelss access nodes' should be 'wireless access nodes'.
[Section II-A] The phrase 'In particularly' should be 'In particular'.
[Section II-B] The textual description in Eq. (8) says 'reprensented' instead of 'represented'.
[Section II-B] Eq. (13) writes 'max I(s; b_hat) = E[log p(s|b_hat)] + const'; this is a variational lower bound approach, and the equality should be stated as 'maximizing a lower bound on I' to be precise. The subsequent claim that this is 'equivalent to minimizing MSE' is standard in the VAE literature, but the bound nature should be acknowledged.
[Section II-B] The regularization in Eq. (9) is described as preventing collapse to error-free solutions (epsilon_n -> 0), but the term (epsilon_n - 0.5)^2 actually pushes epsilon toward 0.5, the maximum-entropy point. The stated purpose is better phrased as encouraging a spread of bit-flipping probabilities rather than merely avoiding zero.
[Section II-B] In the sentence following Eq. (12), 'q_noise' should be 'q_noisy' for consistency with Eq. (10).
[Section III-B] Typo: 'our porposed model' should be 'our proposed model'.
[Section III-C] The sentence 'This further confirms that the best performance is achieved only when both the Interface Attention Module and the SE Block (i.e., the complete IAN) are used' is confusing because the previous sentence says that removing only the Interface Attention Module (retaining SE Block) degrades performance compared to removing both; this seems to imply the SE Block alone hurts. The authors likely mean that the SE Block helps only when combined with the Interface Attention Module, but the text should be rephrased for clarity.
[Section III-A.4] The choice of lambda = 1 in Eq. (15) is not justified or studied. A sensitivity analysis over lambda would strengthen the paper, though it is not essential for the main claim.
[Section III-A.1] The description of model training per dataset is ambiguous: the text says 'we train the model using the training samples from CIFAR10' but later implies separate training for SVHN and ImageNet32. Please clarify which models are trained on which datasets and which results correspond to cross-dataset generalization.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the learned interface is an empirical design validated by ablation, not a derivation that reduces to its inputs.

full rationale

The paper contains no equation-level circularity. The first stage trains the source mapper/demapper and the interface parameters {epsilon_n} by maximizing a variational lower bound on I(s; b_hat) (Eqs. 10-15), using BSC noise as a stand-in for the end-to-end channel. The second stage freezes the source mapper/demapper and the interface, then trains the channel mapper/demapper against the same squared-error reconstruction loss (Eq. 16); the same epsilon values are imported as 1-2epsilon features into the IAN. This is a learned-feature transfer, not a prediction that reduces to a fitted value: the paper nowhere claims to derive reconstruction PSNR from epsilon, and the IAN contribution is tested by an ablation that removes the interface attention module (Section III-C). No load-bearing self-citation appears: references [6], [9]-[13] are external. No uniqueness theorem is invoked to force the interface form. The use of 1-2epsilon as an importance score is a stated design choice, not a renamed known result presented as unification. The asymmetric training SNR between the proposed model (uniform 5-20 dB) and Split DeepJSCC (fixed 20 dB) is a real experimental confound that threatens the headline comparison, but it is a fairness/correctness issue and not a circularity of the derivation, so it does not raise the circularity score under the rubric.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on a standard variational bound, a BSC proxy for the channel, and a transfer assumption that the learned epsilon ordering remains informative for real-channel training. The main fitted quantities are the M interface bit-flip probabilities and the hand-set regularization lambda; no learned values are reported.

free parameters (3)

bit-flipping probabilities {epsilon_n} = not reported
M trainable interface parameters learned in stage 1 to maximize variational MI; used later as bit-importance via 1-2 epsilon. No learned values are reported.
regularization coefficient lambda = 1
Hand-chosen to trade off preventing collapse of epsilon to 0; no sensitivity analysis.
training SNR distribution = uniform 5-20 dB
Chosen design choice for proposed method; baseline uses fixed 20 dB, creating an asymmetric comparison.

assumptions (4)

domain assumption The composition of channel mapper, wireless channel, and channel demapper can be modeled as a bank of independent binary symmetric channels for source code training.
Stage 1 trains source mapper and demapper through BSCs only; this assumes the BSC is a sufficient proxy for the real channel code, which is not proven.
standard math The variational approximation p_phi(s|b_hat) with a Gaussian output reduces MI maximization to MSE minimization.
Standard variational identity used in [10], [11]; valid under Gaussian posterior assumption.
standard math Straight-Through Estimator provides usable gradients through the non-differentiable binary sampling.
Common heuristic; not a guarantee but widely used.
domain assumption The learned 1-2 epsilon importance weights transfer to the real channel mapper and demapper training.
Stage 2 feeds interface-derived importance into the IAN; the paper assumes this ordering remains informative for the actual wireless channel, validated only by ablation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning-Based Interface for Semantic Communication with Bit Importance Awareness." pith.science (2026). https://pith.science/paper/JBRPBWZX

@misc{pith2026250712850,
  author       = {Pith},
  title        = {Pith review of: Learning-Based Interface for Semantic Communication with Bit Importance Awareness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JBRPBWZX}},
  note         = {Machine review of arXiv:2507.12850}
}

read the original abstract

Joint source-channel coding (JSCC) is an effective approach for semantic communication. However, current JSCC methods are difficult to integrate with existing communication network architectures, where application and network providers are typically different entities. Recently, a novel paradigm termed Split DeepJSCC has been under consideration to address this challenge. Split DeepJSCC employs a bit-level interface that enables separate design of source and channel codes, ensuring compatibility with existing communication networks while preserving the advantages of JSCC in terms of semantic fidelity and channel adaptability. In this paper, we propose a learning-based interface design by treating its parameters as trainable, achieving improved end-to-end performance compared to Split DeepJSCC. In particular, the interface enables specification of bit-level importance at the output of the source code. Furthermore, we propose an Importance-Aware Net that utilizes the interface-derived bit importance information, enabling dynamical adaptation to diverse channel bandwidth ratios and time-varying channel conditions. Experimental results show that our method improves performance in wireless image transmission tasks. This work provides a potential solution for realizing semantic communications in existing wireless networks.

Figures

Figures reproduced from arXiv: 2507.12850 by the authors.

**Figure 1.** System model. A. Overall Architecture The overview of our system model is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 3.** Illustration of model training. where MAX represents the maximum possible value of each element in s (e.g., 255 for an 8-bit RGB pixel), and the mean-squared error (MSE) is calculated as MSE(s, sˆ) = ∥s − sˆ∥ 2 2 . Notably, unlike traditional separate source-channel coding (SSCC), our approach does not prioritize accurate reconstruction of the bit sequence ˆb, and instead focuses on image reconstruction, enabling se… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Structure of the attention module. The structure of the IAN is depicted in Fig. 4b. For the interface attention module, we modify the conventional convolutional attention mechanism (shown in Fig. 4a) by directly integrating interface information into latent features. A…

**Figure 5.** Figure 5: PSNR versus SNR over AWGN channel (a) CIFAR10 (b) CIFAR100 (c) SVHN (d) ImageNet32 [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: PSNR versus SNR over Rayleigh fading channel [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: PSNR comparison with/without Importance Aware Net. [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]

**Figure 8.** Figure 8: Examples of visual comparison. C. Ablation Study In order to show the effectiveness of our proposed IAN, we evaluated PSNR performance under two channel conditions (i.e., AWGN and Rayleigh fading) on the CIFAR10 dataset with the CBR fixed at 1/24 using four schemes: 1)…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 8 canonical work pages

[9]

Digital deep joint source-channel coding with blind training for adaptive modulation and power control,

Y . Oh, J. Park, J. Choi, J. Park, and Y .-S. Jeon, “Digital deep joint source-channel coding with blind training for adaptive modulation and power control,” arXiv preprint arXiv:2501.02273 , 2025

arXiv 2025
[13]

Witt: A wireless image transmission transformer for semantic communications,

K. Yang, S. Wang, J. Dai, K. Tan, K. Niu, and P. Zhang, “Witt: A wireless image transmission transformer for semantic communications,” in 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

work page 2023
[1]

C. E. Shannon and W. Weaver, The Mathematical Theory of Communi- cation. University of Illinois press, 1962

work page 1962
[2]

Deep learning enabled semantic communication systems,

H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Transactions on Signal Pro- cessing, vol. 69, pp. 2663–2675, 2021

2021
[3]

Deep joint source- channel coding for wireless image transmission,

E. Bourtsoulatze, D. B. Kurka, and D. G ¨und¨uz, “Deep joint source- channel coding for wireless image transmission,” IEEE Transactions on Cognitive Communications and Networking , vol. 5, no. 3, pp. 567–579, 2019

2019
[4]

Deepwive: Deep-learning-aided wireless video transmission,

T.-Y . Tung and D. G ¨und¨uz, “Deepwive: Deep-learning-aided wireless video transmission,” IEEE Journal on Selected Areas in Communica- tions, vol. 40, no. 9, pp. 2570–2583, 2022

work page 2022
[5]

Semantic communication systems for speech transmission,

Z. Weng and Z. Qin, “Semantic communication systems for speech transmission,” IEEE Journal on Selected Areas in Communications , vol. 39, no. 8, pp. 2434–2444, 2021

2021
[6]

Multi- level reliability interface for semantic communications over wireless networks,

T.-Y . Tung, H. Esfahanizadeh, J. Du, and H. Viswanathan, “Multi- level reliability interface for semantic communications over wireless networks,” IEEE Transactions on Communications , 2025

work page 2025

Show all 17 references

[7]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, and S. Gelly, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

2010 arXiv
[8]

Swin transformer: Hierarchical vision transformer using shifted windows,

Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 012–10 022

2021
[10]

Neural joint source-channel coding,

K. Choi, K. Tatwawadi, A. Grover, T. Weissman, and S. Ermon, “Neural joint source-channel coding,” in International Conference on Machine Learning. PMLR, 2019, pp. 1182–1192

2019
[11]

Joint source-channel coding for channel-adaptive digital semantic communications,

J. Park, Y . Oh, S. Kim, and Y .-S. Jeon, “Joint source-channel coding for channel-adaptive digital semantic communications,” IEEE Transactions on Cognitive Communications and Networking , pp. 75–89, 2024

2024
[12]

Estimating or propagating gradients through stochastic neurons for conditional computation,

Y . Bengio, N. L ´eonard, and A. Courville, “Estimating or propagating gradients through stochastic neurons for conditional computation,” arXiv preprint arXiv:1308.3432, 2013

2013 arXiv
[14]

Squeeze-and-excitation networks,

J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 7132–7141

2018
[15]

Learning multiple layers of features from tiny images,

A. Krizhevsky and G. Hinton, “Learning multiple layers of features from tiny images,” Handbook of Systemic Autoimmune Diseases , vol. 1, no. 4, 2009

2009
[16]

Reading digits in natural images with unsupervised feature learning,

Y . Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y . Ng, “Reading digits in natural images with unsupervised feature learning,” in NIPS Workshop on Deep Learning and Unsupervised Feature Learning , vol. 2011, no. 2. Granada, 2011, p. 4

2011
[17]

A downsampled variant of imagenet as an alternative to the cifar datasets,

P. Chrabaszcz, I. Loshchilov, and F. Hutter, “A downsampled variant of imagenet as an alternative to the cifar datasets,” arXiv preprint arXiv:1707.08819, 2017

2017 arXiv

Pith tools

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

[1] [9]

Digital deep joint source-channel coding with blind training for adaptive modulation and power control,

Y . Oh, J. Park, J. Choi, J. Park, and Y .-S. Jeon, “Digital deep joint source-channel coding with blind training for adaptive modulation and power control,” arXiv preprint arXiv:2501.02273 , 2025

arXiv 2025

[2] [13]

Witt: A wireless image transmission transformer for semantic communications,

K. Yang, S. Wang, J. Dai, K. Tan, K. Niu, and P. Zhang, “Witt: A wireless image transmission transformer for semantic communications,” in 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

work page 2023

[3] [1]

C. E. Shannon and W. Weaver, The Mathematical Theory of Communi- cation. University of Illinois press, 1962

work page 1962

[4] [2]

Deep learning enabled semantic communication systems,

H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Transactions on Signal Pro- cessing, vol. 69, pp. 2663–2675, 2021

2021

[5] [3]

Deep joint source- channel coding for wireless image transmission,

E. Bourtsoulatze, D. B. Kurka, and D. G ¨und¨uz, “Deep joint source- channel coding for wireless image transmission,” IEEE Transactions on Cognitive Communications and Networking , vol. 5, no. 3, pp. 567–579, 2019

2019

[6] [4]

Deepwive: Deep-learning-aided wireless video transmission,

T.-Y . Tung and D. G ¨und¨uz, “Deepwive: Deep-learning-aided wireless video transmission,” IEEE Journal on Selected Areas in Communica- tions, vol. 40, no. 9, pp. 2570–2583, 2022

work page 2022

[7] [5]

Semantic communication systems for speech transmission,

Z. Weng and Z. Qin, “Semantic communication systems for speech transmission,” IEEE Journal on Selected Areas in Communications , vol. 39, no. 8, pp. 2434–2444, 2021

2021

[8] [6]

Multi- level reliability interface for semantic communications over wireless networks,

T.-Y . Tung, H. Esfahanizadeh, J. Du, and H. Viswanathan, “Multi- level reliability interface for semantic communications over wireless networks,” IEEE Transactions on Communications , 2025

work page 2025