An SCMA Receiver for 6G NTN based on Multi-Task Learning

Alessandro Guidotti; Alessandro Vanelli-Coralli; Bruno De Filippo; Carla Amatetti; Riccardo Campana

REVIEW 4 major objections 6 minor 19 references

An SCMA Receiver for 6G NTN based on Multi-Task Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A multi-task CNN receiver decodes overlapping SCMA signals in a LEO satellite uplink with about 3.25 dB less energy per bit than the Log-MPA benchmark at 10% block error rate.

desk verdict A solid link-level simulation showing a multi-task CNN SCMA receiver gains 3–3.75 dB over Log-MPA in a LEO uplink, with the ideal-CSI caveat honestly stated but worth an explicit robustness check. read the letter →

arxiv 2506.05111 v1 pith:KXJK4ACA submitted 2025-06-05 eess.SP

Bruno De Filippo , Carla Amatetti , Riccardo Campana , Alessandro Guidotti , Alessandro Vanelli-Coralli This is my paper

classification eess.SP

keywords SCMAsparsecodemultipleaccessmulti-tasklearningdeepreceivernon-terrestrialnetworksLEOsatellite6Gblockerrorrate

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 proposes a convolutional neural network as the receiver for sparse code multiple access (SCMA) in a low-Earth-orbit satellite uplink, where several ground devices transmit on the same subcarriers and the satellite hosts an edge-computing server. The authors train the network with multi-task learning, giving it one shared feature extractor plus a separate output chain for each user, so that it estimates log-likelihood ratios for all users from one received symbol grid. In link-level simulations, the network reaches the 10% block-error-rate target at about 3.25 dB lower energy per bit than the standard Log-MPA algorithm at code rate 0.588, and 0.5 dB lower at code rate 0.188. This lets the link operate at a higher code rate and lower signal-to-noise ratio, increasing the aggregated throughput at the satellite. The practical motivation is that AI-based reception could reduce ground-terminal complexity and power while keeping the satellite payload feasible with hardware acceleration.

What carries the argument

The mechanism is a 1-D convolutional network with eight Conv1D layers (256 kernels of length 3) shared across users, followed by max pooling and two Conv1D layers, then J parallel dense chains (three 256-neuron layers plus an m-neuron output) that specialize per user. The input is a $(K, 2(J+1))$ real matrix per symbol time: received symbols and known channel coefficients arranged so that the SCMA codebook sparsity pattern is embedded, which lets the convolution layers learn the interference structure. Training uses binary cross-entropy with logits, so the pre-activation output is directly the log-likelihood ratio for each coded bit; this is what makes the CNN output usable by the LDPC decoder without extra conversion.

What would settle it

Re-run the same Monte Carlo link-level simulation using channel estimates from a pilot-based scheme (for example, the 10% pilot overhead approach cited in the paper) instead of exact coefficients, and compare the $E_b/N_0$ needed for 10% BLER; if the CNN's gain over Log-MPA drops below roughly 1 dB, the central claim is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that a CNN trained with multi-task learning can replace the iterative Log-MPA receiver for SCMA over a LEO NTN link without sacrificing error performance, and in fact improving it at low $E_b/N_0$. For a six-user, four-subcarrier SCMA codebook (150% overloading) with LDPC coding and a single-tap line-of-sight channel, the network achieves 10% BLER at $-2.75$ dB $E_b/N_0$ for $R_c = 0.588$ and $-5.25$ dB for $R_c = 0.188$; Log-MPA needs 3.25 dB and 0.5 dB more, respectively, and the advantage translates into higher aggregated theoretical throughput. The paper also reports an error floor above $3\cdot10^{-3}$ at high code rate and $5\cdot10^{-4}$ at low code rate rather than a waterfall curve, attributing this to residual approximation error of the neural model.

Load-bearing premise

The load-bearing premise is the assumption of perfect channel knowledge at the receiver; if real pilot-based channel estimation is significantly worse, especially under LEO Doppler, the reported $E_b/N_0$ gains may not survive.

Editorial extensions

If this is right

The receiver lets a higher code rate (0.588) be used at low $E_b/N_0$, where the benchmark can only support a lower rate, yielding higher aggregated throughput at the satellite.
The same ground devices can reduce transmit power or extend battery life while keeping the 10% BLER target, since the required $E_b/N_0$ drops.
The CNN adds no complexity at the ground terminals; all processing moves to the satellite payload, where hardware accelerators and model compression can make the 7.9M MAC operations feasible.
Because the output is LLRs, the network plugs directly into existing LDPC decoding chains without changes to the channel decoder.
At high $E_b/N_0$ the receiver shows an error floor above $3\cdot10^{-3}$ (high rate) or $5\cdot10^{-4}$ (low rate), so it is best suited for non-critical traffic or as part of a hybrid receiver.

Reading between the lines

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

If channel estimates from realistic pilots replace the perfect channel knowledge assumed in Section III, the reported $E_b/N_0$ gain may shrink; the paper itself flags imperfect CSI over non-orthogonal pilots as an open issue in LEO SCMA.
The multi-task structure could be reused for a different SCMA codebook or for grant-free access by retraining only the per-user output chains, something the paper does not demonstrate.
Because the simulated channel is a single-tap line-of-sight model, the gain should be re-tested under multipath or Doppler estimation error before deployment; a terrestrial variant of the same receiver is a natural next experiment.

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

You should know two things about this one. First, it is a competently executed link-level simulation, not a hand-wavy ML paper: LDPC-coded, 3GPP 38.811-derived channel, Monte Carlo with 10^4 iterations, and a genuine benchmark (Log-MPA). Second, the headline gain is real only under perfect channel knowledge; the authors say so in Section III and again in Section VI, but the paper would be stronger if it quantified how much CSI error eats into that gain.

What is actually new: a multi-task CNN SCMA receiver evaluated end-to-end in a channel-coded LEO NTN uplink, which the cited literature does not contain. The architecture is a combination of known pieces (DeepRx-style preprocessing, CNN detector, task-specific heads), but the combination is applied to a new scenario with a credible simulation setup. The observed 3.25 dB gain over Log-MPA at 10% BLER for Rc=0.588 is plausible, and the much smaller 0.5 dB gain at Rc=0.188 is consistent with LDPC correcting most demodulation errors. The complexity discussion (7.9M MAC vs 23k) is honest about the 340x gap and does not oversell it.

Soft spots, in proportion. The three different numbers for the gain (abstract 3.5 dB, Section IV 3.25 dB, conclusion up to 3.75 dB) are sloppy and should be reconciled with the actual curve. No code or data is released, and there are no error bars; for an ML paper, that limits reproducibility, though the training details are described in enough detail to reimplement. The benchmark is approximate Log-MPA, not full MPA, so part of the gain is simply closing the gap to a better detector; that is acceptable if stated, but a full-MPA reference would make the claim cleaner. The bigger issue is the one the stress-test flags: the receiver assumes perfect CSI, and the gain could shrink or vanish under Doppler estimation error. I would not call that a fatal flaw, because the paper is explicit about the assumption and flags it as future work, but it is the load-bearing limitation. A single simulation with a channel estimator with 1 dB error would settle it.

Citation pattern is fine; prior SCMA-NTN work is referenced, and self-citations are limited and relevant.

Bottom line: this is a useful engineering result for SCMA-in-NTN researchers, and it deserves a serious referee. It needs revision: reconcile the gain numbers, add error bars or at least discuss Monte Carlo variance, and ideally add an imperfect-CSI point. I would not desk-reject it, and I would be happy to see it published after those fixes.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes a multi-task learning (MTL) convolutional neural network (CNN) to replace the iterative message passing algorithm (MPA) as the SCMA receiver in a LEO non-terrestrial network (NTN) uplink. The receiver takes as input the received frequency-domain grid and perfect channel coefficients and outputs bit LLRs for each user, which are then decoded by a 5G LDPC decoder. The authors evaluate the CNN in a link-level simulator under AWGN and a single-tap line-of-sight channel model, comparing BLER and aggregated throughput against Log-MPA with 10 iterations. They report that the CNN reaches 10% BLER at an Eb/N0 about 3.25 dB lower than Log-MPA at code rate 0.588 and 0.5 dB lower at 0.188, with corresponding throughput improvements. The paper also estimates the CNN's computational complexity in terms of MAC operations and discusses hardware accelerator feasibility.

Significance. The strength of the paper is that it provides a concrete, reproducible-by-description architecture and an end-to-end evaluation that includes channel coding, which is more realistic than many prior SCMA deep-learning studies. The MTL structure that specializes per codebook is a sensible design, and the complexity analysis with a specific MAC count and a hardware example is useful. However, the headline gain is contingent on ideal channel state information and on the choice of Log-MPA as the benchmark, and the reported gain values are numerically inconsistent across sections. The lack of error bars on the Monte Carlo results further weakens the quantitative claim. If the gain survives imperfect CSI, the result would be significant for 6G NTN research; at present the evidence is not sufficient.

major comments (4)

[Abstract; Sec. IV; Sec. VI] The reported gain at the 10% BLER target is inconsistent: the abstract states 3.5 dB, Sec. IV states 3.25 dB for Rc=0.588 and 0.5 dB for Rc=0.188, and Sec. VI states "up to 3.75 dB." No explanation reconciles these values. Because the magnitude of the Eb/N0 gain is the paper's central quantitative claim, the authors must provide one consistent set of numbers, with the exact operating points, or explain the provenance of each figure.
[Sec. III and Eq. (1)-(2); Sec. VI] The entire evaluation assumes perfect CSI: Sec. III states the channel coefficients "are assumed to be known at the receiver," and Eq. (1) is a single-tap LoS model without Doppler estimation error. Sec. VI acknowledges that SCMA is limited by channel estimation accuracy over non-orthogonal pilots, exacerbated by Doppler in LEO. The reported gains are therefore conditioned on ideal CSI; since the gain at Rc=0.188 is only 0.5 dB, it may be erased by a realistic CSI error. This is a load-bearing limitation, not a future-work item. The authors should provide a sensitivity analysis with imperfect CSI (e.g., a pilot-based channel estimator or a fixed estimation error) or explicitly restrict the claim to ideal CSI with a quantitative argument for why the gain should persist.
[Sec. IV] The BLER results are presented as single curves with no error bars or confidence intervals, despite N_MC = 10^4 Monte Carlo iterations. At a BLER of 0.1, the standard error is on the order of 0.003, but the reported Eb/N0 gains (3.25 dB vs. 0.5 dB) are read from interpolated curves; a small statistical fluctuation could affect the low-rate comparison. The authors should provide confidence intervals, error bars, or at least a statistical-accuracy discussion for the operating points at which the gains are reported.
[Sec. II and Sec. IV] The benchmark is Log-MPA with Niter=10, which is an approximate version of MPA. The paper does not discuss how the reported gain would change if the benchmark were the full MPA or Log-MPA with a larger number of iterations. Since the CNN's gain is defined relative to this specific approximate baseline, at least a sensitivity check on Niter (e.g., 5, 10, 20) is needed to establish that the gain is a property of the CNN rather than an artifact of the baseline's approximation error.

minor comments (6)

[Sec. II] The notation "N P RB SC" is likely a formatting error; the intended meaning is not immediately clear and should be clarified.
[Sec. IV] The phrase "excellent performance" is subjective; a quantitative statement would be more appropriate.
[Sec. IV, Eq. (8)] Equation (8) uses T_TB, but its definition is only given later in the same paragraph; define it with the equation or before it.
[Sec. V] The complexity comparison only reports MAC counts; memory footprint and energy per inference would be more informative for the on-board feasibility claim.
[References] Reference [15] is a bare URL without a publication date or a stable identifier; it should be completed.
[Fig. 3] The text describes the architecture in detail, but the figure itself has small fonts and the task-specific layers are only distinguished by light blue shading, which may be difficult to perceive in print.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CNN BLER gain is an empirical measurement against an external Log-MPA benchmark, not a fitted quantity relabeled as a prediction.

full rationale

The paper's central claim is the measured BLER gap between the proposed CNN and Log-MPA in a Monte Carlo link-level simulation. The CNN is trained to minimize the binary cross-entropy loss (Eqs. 6-7) on simulated SCMA examples, and the reported Eb/N0 operating points are read from the resulting BLER curves in Figure 5, not derived from the training objective. Log-MPA is an independently implemented benchmark with a fixed number of iterations. The only self-citation, reference [1] by Guidotti et al., is used to motivate the importance of NTNs in the introduction and is not load-bearing. The ideal-CSI assumption stated in Section III and the Section VI limitation regarding channel estimation over non-orthogonal pilots and Doppler shift weaken external validity, but they are not circular: the comparison is internally consistent because Log-MPA receives the same perfect CSI. No equation, parameter, or theorem in the paper reduces by construction to the measured gain, and no fitted value is presented as an independent prediction. The evaluation is standard supervised testing on independently generated data, so the derivation chain is self-contained.

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

The central claim depends on a simulation with perfect CSI, a single-tap LoS channel, and manually chosen CNN and Log-MPA configurations. No new physical entities are introduced; the CNN weights are fitted to simulated data, so the main free parameters are the architecture and training choices rather than a derived model.

free parameters (4)

CNN architecture sizes = 8 Conv1D layers of 256 kernels, 2 Conv1D layers of 512 kernels, 3 dense layers of 256 per task head
Architecture chosen by hand with no ablation or sensitivity analysis, so the reported gain is tied to this configuration.
Training hyperparameters = Adam lr=1e-3, decay 10x after 50 epochs without improvement, early stopping 200 epochs, 256 minibatches of 3000 examples
Selected by hand; no optimization or robustness study is reported.
Log-MPA benchmark iterations = 10
The comparison gain depends on this chosen iteration count; a different Log-MPA configuration would change the reported gap.
Training Eb/N0 distribution = not specified
The paper does not report the range or distribution of Eb/N0 used to generate training examples, so it is unknown whether evaluation across -15 to 10 dB is in-distribution or extrapolation.

assumptions (4)

domain assumption Perfect channel state information at the receiver
Section III assumes known channel coefficients; the conclusion acknowledges this is unrealistic for non-orthogonal pilots in LEO.
domain assumption Single-tap LoS channel without Doppler estimation error
Eq. (1) models each link as one tap with free-space path loss and phase; Doppler is not included in the receiver model, although LEO links have large Doppler and Doppler rate.
domain assumption Representativeness of the 3GPP NTN channel model [16]
The results depend on the specific 38.811 simulation assumptions; the paper provides no sensitivity to other channel conditions.
ad hoc to paper Log-MPA is an adequate benchmark
Only Log-MPA with 10 iterations is compared; no full MPA or alternative DL receiver baseline is included.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An SCMA Receiver for 6G NTN based on Multi-Task Learning." pith.science (2026). https://pith.science/paper/KXJK4ACA

@misc{pith2026250605111,
  author       = {Pith},
  title        = {Pith review of: An SCMA Receiver for 6G NTN based on Multi-Task Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KXJK4ACA}},
  note         = {Machine review of arXiv:2506.05111}
}

read the original abstract

Future 6G networks are envisioned to enhance the user experience in a multitude of different ways. The unification of existing terrestrial networks with non-terrestrial network (NTN) components will provide users with ubiquitous connectivity. Multi-access edge computing (MEC) will enable low-latency services, with computations performed closer to the end users, and distributed learning paradigms. Advanced multiple access schemes, such as sparse code multiple access (SCMA), can be employed to efficiently move data from edge nodes to spaceborne MEC servers. However, the non-orthogonal nature of SCMA results in interference, limiting the effectiveness of traditional SCMA receivers. Hence, NTN links should be protected with robust channel codes, significantly reducing the uplink throughput. Thus, we investigate the application of artificial intelligence (AI) to SCMA receivers for 6G NTNs. We train an AI model with multi-task learning to optimally separate and receive superimposed SCMA signals. Through link level simulations, we evaluate the block error rate (BLER) and the aggregated theoretical throughput achieved by the AI model as a function of the received energy per bit over noise power spectral density ratio (Eb/N0). We show that the proposed receiver achieves a target 10% BLER with 3.5dB lower Eb/N0 with respect to the benchmark algorithm. We conclude the assessment discussing the complexity-related challenges to the implementation of the AI model on board of a low earth orbit satellite.

Figures

Figures reproduced from arXiv: 2506.05111 by the authors.

**Figure 2.** SCMA factor graph. resources. The data payload of the i-th GD is mapped to one or more 5G transport blocks (TBs) bi of size NT BS bits. Each TB is encoded with a low density parity check (LDPC) channel encoder with code rate Rc, resulting in a vector of Nbits coded bits b (c) i . An SCMA encoder and modulator is implemented based on the chosen SCMA codebook Γi(·), mapping each group of m coded bits in b (c) i to one… view at source ↗

**Figure 3.** Diagram of the proposed CNN-based SCMA receiver. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** NN input after the pre-processing reshaping step. [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: BLER as a function of the Eb/N0. gap at high code rate is twofold: firstly, CNNs typically provide good performance in pattern recognition tasks in presence of noise; secondly, Log-MPA introduces approximations to MPA to lower its computational complexity, which in tu…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 11 canonical work pages

[1]

Guidotti et al., ”Role and Evolution of Non-Terrestrial Networks Toward 6G Systems,” in IEEE Access, vol

A. Guidotti et al., ”Role and Evolution of Non-Terrestrial Networks Toward 6G Systems,” in IEEE Access, vol. 12, pp. 55945-55963, 2024, doi: 10.1109/ACCESS.2024.3389459

arXiv 2024
[2]

Y . Jing, J. Wang, C. Jiang and Y . Zhan, ”Satellite MEC with Feder- ated Learning: Architectures, Technologies and Challenges,” in IEEE Network, vol. 36, no. 5, pp. 106-112, September/October 2022, doi: 10.1109/MNET.001.2200202

work page doi:10.1109/mnet.001.2200202 2022
[3]

Zhang, Q

J. Zhang, Q. He, Z. Yu, B. Bai and M. Zhu, ”Low-Complexity Coherent Iterative Receiver for SCMA-Based LEO Satellite Commu- nications,” 2019 IEEE Global Communications Conference (GLOBE- COM), Waikoloa, HI, USA, 2019, pp. 1-6, doi: 10.1109/GLOBE- COM38437.2019.9014189

work page arXiv 2019
[4]

Y . Hao, Z. Song, Z. Zheng, Q. Zhang and Z. Miao, ”Joint Communi- cation, Computing, and Caching Resource Allocation in LEO Satellite MEC Networks,” in IEEE Access, vol. 11, pp. 6708-6716, 2023, doi: 10.1109/ACCESS.2023.3237701

arXiv 2023
[5]

Y . Wang, X. Ding, P. Zhang, X. Xie and G. Zhang, ”Joint CRDSA and SCMA for Satellite Internet of Things with Codebook Collisions,” 2023 IEEE Globecom Workshops (GC Wkshps), Kuala Lumpur, Malaysia, 2023, pp. 129-134, doi: 10.1109/GCWkshps58843.2023.10464815

arXiv 2023
[6]

D. Li, X. Liu, Z. Yin, N. Cheng and J. Liu, ”CWGAN-Based Channel Modeling of Convolutional Autoencoder-Aided SCMA for Satellite- Terrestrial Communication,” in IEEE Internet of Things Journal (in press), doi: 10.1109/JIOT.2024.3425470

arXiv
[7]

C. Han, H. Zhao, Z. Chen and F. Wang, ”Sparse Neural Network for Detection and Decoding of Non-Binary Polar-Coded SCMA,” in IEEE Transactions on Wireless Communications, vol. 22, no. 7, pp. 4475- 4488, July 2023, doi: 10.1109/TWC.2022.3225841

work page arXiv 2023
[8]

Cheng, Y

H. Cheng, Y . Xia, Y . Huang, Z. Lu and L. Yang, ”Deep Neural Network Aided Low-Complexity MPA Receivers for Uplink SCMA Systems,” in IEEE Transactions on Vehicular Technology, vol. 70, no. 9, pp. 9050- 9062, Sept. 2021, doi: 10.1109/TVT.2021.3099640

work page arXiv 2021

Show all 19 references

[9]

Thushan, S

S. Thushan, S. Ali, N. H. Mahmood, N. Rajatheva and M. Latva- Aho, ”Deep Learning-Based Blind Multiple User Detection for Grant- Free SCMA and MUSA Systems,” in IEEE Transactions on Machine Learning in Communications and Networking, vol. 1, pp. 61-77, 2023, doi: 10.1109/TMLCN....

2023
[10]

M. Kim, N. -I. Kim, W. Lee and D. -H. Cho, ”Deep Learning-Aided SCMA,” in IEEE Communications Letters, vol. 22, no. 4, pp. 720-723, April 2018, doi: 10.1109/LCOMM.2018.2792019

2018
[11]

Miuccio, D

L. Miuccio, D. Panno and S. Riolo, ”A Flexible Encoding/Decoding Procedure for 6G SCMA Wireless Networks via Adversarial Ma- chine Learning Techniques,” in IEEE Transactions on Vehicular Technology, vol. 72, no. 3, pp. 3288-3303, March 2023, doi: 10.1109/TVT.2022.3216028

2023
[12]

Jiang, D

F. Jiang, D. -W. Chang, S. Ma, Y . -J. Hu and Y . -H. Xu, ”A Resid- ual Learning-Aided Convolutional Autoencoder for SCMA,” in IEEE Communications Letters, vol. 27, no. 5, pp. 1337-1341, May 2023, doi: 10.1109/LCOMM.2023.3260881

2023
[13]

J. Liu, G. Wu, S. Li and O. Tirkkonen, ”On Fixed-Point Imple- mentation of Log-MPA for SCMA Signals,” in IEEE Wireless Com- munications Letters, vol. 5, no. 3, pp. 324-327, June 2016, doi: 10.1109/LWC.2016.2554557

2016
[14]

Honkala, D

M. Honkala, D. Korpi and J. M. J. Huttunen, ”DeepRx: Fully Con- volutional Deep Learning Receiver,” in IEEE Transactions on Wire- less Communications, vol. 20, no. 6, pp. 3925-3940, June 2021, doi: 10.1109/TWC.2021.3054520

2021
[15]

”1st 5G Algorithm Innovation Competition-ENVl.0-SCMA”, Available at: http://www.innovateasia.com/5g/en/gp2.html
[16]

”38.811 - Study on New Radio (NR) to support non-terrestrial net- works”, 2020

2020
[17]

TensorFlow: Large-scale machine learning on hetero- geneous systems

M. Abadi et al, “TensorFlow: Large-scale machine learning on hetero- geneous systems”, 2015. Software available from tensorflow.org

2015
[18]

J. Guo, J. Wang, C. -K. Wen, S. Jin and G. Y . Li, ”Compression and Acceleration of Neural Networks for Communications,” in IEEE Wireless Communications, vol. 27, no. 4, pp. 110-117, August 2020, doi: 10.1109/MWC.001.1900473

2020 doi
[19]

Lamberti et al., ”Distilling Tiny and Ultrafast Deep Neural Networks for Autonomous Navigation on Nano-UA Vs,” in IEEE Internet of Things Journal, vol

L. Lamberti et al., ”Distilling Tiny and Ultrafast Deep Neural Networks for Autonomous Navigation on Nano-UA Vs,” in IEEE Internet of Things Journal, vol. 11, no. 20, pp. 33269-33281, 15 Oct.15, 2024, doi: 10.1109/JIOT.2024.3431913

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Guidotti et al., ”Role and Evolution of Non-Terrestrial Networks Toward 6G Systems,” in IEEE Access, vol

A. Guidotti et al., ”Role and Evolution of Non-Terrestrial Networks Toward 6G Systems,” in IEEE Access, vol. 12, pp. 55945-55963, 2024, doi: 10.1109/ACCESS.2024.3389459

arXiv 2024

[2] [2]

Y . Jing, J. Wang, C. Jiang and Y . Zhan, ”Satellite MEC with Feder- ated Learning: Architectures, Technologies and Challenges,” in IEEE Network, vol. 36, no. 5, pp. 106-112, September/October 2022, doi: 10.1109/MNET.001.2200202

work page doi:10.1109/mnet.001.2200202 2022

[3] [3]

Zhang, Q

J. Zhang, Q. He, Z. Yu, B. Bai and M. Zhu, ”Low-Complexity Coherent Iterative Receiver for SCMA-Based LEO Satellite Commu- nications,” 2019 IEEE Global Communications Conference (GLOBE- COM), Waikoloa, HI, USA, 2019, pp. 1-6, doi: 10.1109/GLOBE- COM38437.2019.9014189

work page arXiv 2019

[4] [4]

Y . Hao, Z. Song, Z. Zheng, Q. Zhang and Z. Miao, ”Joint Communi- cation, Computing, and Caching Resource Allocation in LEO Satellite MEC Networks,” in IEEE Access, vol. 11, pp. 6708-6716, 2023, doi: 10.1109/ACCESS.2023.3237701

arXiv 2023

[5] [5]

Y . Wang, X. Ding, P. Zhang, X. Xie and G. Zhang, ”Joint CRDSA and SCMA for Satellite Internet of Things with Codebook Collisions,” 2023 IEEE Globecom Workshops (GC Wkshps), Kuala Lumpur, Malaysia, 2023, pp. 129-134, doi: 10.1109/GCWkshps58843.2023.10464815

arXiv 2023

[6] [6]

D. Li, X. Liu, Z. Yin, N. Cheng and J. Liu, ”CWGAN-Based Channel Modeling of Convolutional Autoencoder-Aided SCMA for Satellite- Terrestrial Communication,” in IEEE Internet of Things Journal (in press), doi: 10.1109/JIOT.2024.3425470

arXiv

[7] [7]

C. Han, H. Zhao, Z. Chen and F. Wang, ”Sparse Neural Network for Detection and Decoding of Non-Binary Polar-Coded SCMA,” in IEEE Transactions on Wireless Communications, vol. 22, no. 7, pp. 4475- 4488, July 2023, doi: 10.1109/TWC.2022.3225841

work page arXiv 2023

[8] [8]

Cheng, Y

H. Cheng, Y . Xia, Y . Huang, Z. Lu and L. Yang, ”Deep Neural Network Aided Low-Complexity MPA Receivers for Uplink SCMA Systems,” in IEEE Transactions on Vehicular Technology, vol. 70, no. 9, pp. 9050- 9062, Sept. 2021, doi: 10.1109/TVT.2021.3099640

work page arXiv 2021