Differential Transformer-driven 6G Physical Layer for Collaborative Perception Enhancement

Mohamed Ababsa; Osama Saleem; Soheyb Ribouh

REVIEW 3 major objections 5 minor 25 references

Differential Transformer-driven 6G Physical Layer for Collaborative Perception Enhancement

T0 review · 3 major / 5 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 The paper claims that a Differential Transformer neural receiver, built on attention maps formed by subtracting two softmax maps, beats CNN and conventional baselines on bit error rate over standardized Urban Macro and CDL channels, and…

desk verdict BER gains for a Diff-Transformer receiver over DeepRx look real; the collaborative-perception AP numbers rest on an undocumented encoding pipeline and should not be taken at face value. read the letter →

arxiv 2506.20597 v1 pith:BHFAUC6B submitted 2025-06-25 eess.SP

Soheyb Ribouh , Osama Saleem , Mohamed Ababsa This is my paper

classification eess.SP

keywords 6GdifferentialattentionneuralreceivercollaborativeperceptionV2XcommunicationOFDMbiterrorrateconnectedautonomousvehicles

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

Connected autonomous vehicles can only cooperate if the wireless link accurately delivers LiDAR and camera data, so this paper treats the physical layer as a learning problem. It proposes a neural receiver built from Differential Transformer blocks, where attention is computed by subtracting two softmax maps, and claims this architecture replaces channel estimation, equalization, and demapping in an orthogonal frequency-division multiplexing (OFDM) link. Tested on standardized Urban Macro and CDL channels, the receiver is reported to beat a CNN-based receiver and a conventional baseline by wide BER margins, reaching a target BER of $10^{-4}$ at 5.75 dB. The same communication improvement is then shown to propagate through a four-agent heterogeneous fusion setup, where the proposed receiver achieves AP30/AP50/AP70 of 0.89/0.89/0.84. The reader should care because it is a concrete claim that a noise-suppressing attention mechanism can improve both the received bits and the downstream autonomous-driving decisions built on them.

What carries the argument

The load-bearing object is the differential attention map, $\operatorname{DiffAttn} = (\operatorname{softmax}(Q_1K_1^\top/\sqrt{d}) - \operatorname{softmax}(Q_2K_2^\top/\sqrt{d}))V$, where subtracting the two soft attention maps is intended to cancel noisy correlations and keep only signal-consistent context. Around it the receiver stacks an input dense layer, four Differential Transformer blocks with four heads, and an output dense layer that emits log-likelihood ratios (LLRs), the soft bit estimates a decoder consumes, which are passed to an LDPC belief-propagation decoder. This single network replaces the classical chain of channel estimation, equalization, and demapping, and the paper's argument is that the subtraction mechanism is what produces the reported BER and perception gains.

What would settle it

Run the fusion model on bit-exact, untransmitted sensor data and measure its AP; if that bit-exact AP is not substantially above the values in Table I, then the reported receiver-to-receiver AP gaps are not isolating physical-layer performance.

Watch

Extended reading notes

Core claim

The paper's central claim is that differential attention, an attention map formed by subtracting two independent softmax query-key maps, is a better inductive bias for an end-to-end neural OFDM receiver than the convolutional architecture of the CNN baseline. On the Urban Macro channel at 0–60 km/h it reports BER $2\times10^{-4}$ at 6 dB versus $1.2\times10^{-3}$ for the CNN receiver and $10^{-1}$ for the classical baseline; at 60–120 km/h it converges to near-zero BER at 6.25 dB against 8.25 dB and about 20 dB for the two baselines. On an unseen CDL channel it keeps a 0.5 dB advantage over the CNN receiver and a 5 dB advantage over the baseline. The authors then report that these gains transfer to collaborative perception: after four agents' reconstructed LiDAR and camera data are fused, their receiver achieves average precision 0.89 at AP30/AP50 and 0.84 at AP70, versus 0.73/0.73/0.61 for the CNN receiver and 0.00 for the conventional receiver.

Load-bearing premise

The load-bearing premise is that each vehicle's LiDAR and camera data can be turned into bits, sent over the OFDM link, and rebuilt at the receiver without losing anything the fusion model uses, yet the paper never describes that encoding, compression, or reconstruction pipeline.

Editorial extensions

If this is right

At SNR 6 dB over the Urban Macro channel, the proposed receiver reaches a BER of about $2\times10^{-4}$, roughly six times better than the CNN receiver and about three orders of magnitude better than the baseline.
To reach a target BER of $10^{-4}$, the proposed receiver needs 5.75 dB, versus 7.5 dB for the CNN receiver and over 10 dB for the conventional receiver.
Under high mobility (60–120 km/h), the proposed receiver reaches near-zero BER at 6.25 dB, a gain of 2.25 dB over the CNN receiver and 13.75 dB over the baseline, and it keeps a 0.5 dB advantage on the CDL channel.
With four heterogeneous agents, the receiver yields AP30/AP50/AP70 of 0.89/0.89/0.84, compared with 0.73/0.73/0.61 for the CNN receiver and 0.00 for the baseline, implying that communication errors were the dominant cause of perception failure in the conventional setup.
If the architecture is as noise-robust as reported, differential attention is a drop-in component for neural receivers in high-mobility V2X links, since the same receiver trained on the Urban Macro channel also performs well on the CDL channel.

Reading between the lines

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

Our inference: the perception comparison needs a bit-exact upper bound; without the AP of the fusion model fed pristine, untransmitted sensor data, the reported 0.84 cannot be distinguished from what perfect communication would already achieve.
Our inference: the 0.00 AP for the conventional receiver is so extreme that it points to a breakdown in the serialization or reconstruction pipeline, not merely worse BER; a fair extension would corrupt received bits at controlled rates and measure AP to isolate communication losses from perception losses.
Our inference: differential attention's noise rejection could transfer to mmWave beam selection or integrated sensing and communication, where the same subtracted-attention mechanism would suppress interference rather than additive channel noise.

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

Two things to know before reading this one. The physical-layer comparison is a legitimate data point: a Differential Transformer receiver beats a retrained DeepRx by 1.75–2.25 dB on 3GPP UMa and by about 0.5 dB on CDL, with the conventional baseline far behind. That part is coherent and worth a look. The collaborative-perception half, the other half of the title, is not supported as written. Table I depends on a sensor-to-bit encoding, compression, and reconstruction pipeline that is never described. Without it, the AP numbers are unfalsifiable.

What is new: applying differential attention from Ye et al. to an end-to-end neural receiver for V2X, evaluated in Sionna. That is a modest architectural swap within the authors' own receiver line, but the evaluation is fair: DeepRx is retrained on the same data and channel, and the gains are non-trivial. The paper also does well to test two mobility ranges in UMa plus a CDL profile and to report the SNR needed to hit a target BER.

The soft spots are real. The stress-test note is correct: 128 subcarriers, 14 OFDM symbols, and 64-QAM give roughly 10,752 coded bits per frame, and nothing in the paper explains how four agents' LiDAR and camera data can be serialized into that budget, nor how the reconstructed data are fed to HEAL. The Baseline AP of 0.00 is implausible on its face—a conventional receiver that decodes anything should not produce zero detections at every IoU threshold unless the source coding collapses, and that source coding is exactly what is missing. There are also no error bars and no ablations, and no code or data are provided. These are not minor omissions: they make the central perception claim untestable. The physical-layer claim, by contrast, is reproducible in principle with standard Sionna training.

Citation-wise, the self-citations are used appropriately to position the work, and the differential-attention reference is correct. No red flag there.

Bottom line: the BER result is a reasonable incremental contribution to neural receivers; the perception result needs to be redone with the actual data pipeline specified—source coding, bit allocation, frame repetition, reconstruction, and a non-zero baseline. A serious referee could push this into a solid paper. Send it to review, but make sure the reviewers demand the missing pipeline and the artifacts.

Referee Report

3 major / 5 minor

Summary. The manuscript presents a Differential Transformer-based deep learning receiver for OFDM-based V2X links. The receiver processes received complex samples to produce log-likelihood ratios that are then passed to an LDPC decoder, and it is compared against the CNN-based DeepRx receiver and an unnamed conventional baseline in terms of bit error rate over 3GPP UMa and CDL channels. The same receiver is then embedded in a collaborative perception experiment using the HEAL framework on OPV2V data from four heterogeneous agents, with the paper reporting average precision values up to 0.89 at IoU 30/50 and 0.84 at IoU 70. The central claim is that differential attention yields communication gains that translate into improved multi-agent perception accuracy.

Significance. If substantiated, the physical-layer result would be a useful empirical contribution: the paper is among the first to apply differential attention to an end-to-end neural receiver, and it uses standard tools (Sionna, 3GPP channel models, OPV2V, HEAL) that make the comparison meaningful. The comparison against a retrained DeepRx under identical conditions is appropriate, and the reported BER gains are internally coherent. However, the perception-layer half of the claim is not currently supported by the described methodology: the sensor-to-bit encoding and reconstruction pipeline is missing, so the reported AP values cannot be attributed to the proposed receiver. The paper therefore contains a plausible physical-layer study and an unfalsifiable perception study; the latter must be fixed before the central claim can be accepted.

major comments (3)

[II.A, III.B, Table I] The collaborative-perception claim is not reproducible as stated. Section II.A only says that captured LiDAR/camera data are converted into a binary bitstream, and Section III.B says that received data are reconstructed using the proposed Diff-Transformer Neural Receiver, but the paper never specifies the source coding, compression, per-agent bit allocation, frame repetition, or reconstruction algorithm. Given the physical-layer parameters in Section III.A (128 subcarriers, 14 OFDM symbols, 64-QAM), one OFDM frame carries only 10,752 coded bits, which is far below the raw data volume of OPV2V LiDAR and camera streams from four agents. Without knowing how many frames are used and how the receiver maps LLRs back to images and point clouds, the AP values in Table I cannot be reproduced or attributed to receiver quality. Please describe the complete encoding-transmission-reconstruction pipeline, including the operating SNR/BER point, and re-report Table I under this protocol.
[IV.A, Table I] The 'Baseline Receiver' used in Table I is never identified, and its reported AP of exactly 0.00 at all three IoU thresholds is a red flag: even a conventional receiver with a BER of 1e-1 at low SNR would be expected to recover some objects at the lenient AP30 threshold. The paper gives no explanation for this complete failure. Specify the baseline algorithm, its operating SNR, and the detection pipeline; if the number of evaluated frames is small, exact-zero AP may be an artifact of sample size.
[II.B, Eqs. (5)-(8)] The architecture definition is dimensionally inconsistent. Eq. (5) takes V in R^{N×2d} while the queries and keys are in R^{N×d}, so the attention output is N×2d, but Eq. (7) uses W_l in R^{dmodel×dattn} and Eq. (8) uses W_LLR in R^{N×dh}; these shapes do not compose as written. The original differential transformer uses a shared value projection of dimension d, and this also matters for the number of parameters. Please provide a consistent dimension table or release code so the exact architecture is unambiguous.

minor comments (5)

[Throughout] The text contains typos ('bellows', 'decried', 'Diff-Transformerr', 'CA V', 'as decried bellows') that should be corrected in a revision.
[IV.A] The CDL experiment in Fig. 4 does not state which CDL profile is used (e.g., CDL-D or CDL-E) nor the delay/Doppler parameters; please specify so the result is reproducible.
[III.A] The paper reports training on 12 million samples but does not describe the validation/test split, the SNR sampling distribution, or how many independent channel realizations are averaged for each BER point; please add these details and, ideally, error bars.
[III.B] No operating SNR or BER is reported for the perception experiment, so it is unclear whether Table I corresponds to a favorable or unfavorable regime; state the SNR used.
[II] Fig. 1 is referenced as 'shown in figure 1' but the caption and callout are not consistent; place the figure reference at first mention in the text and make sure the notation (Q1, Q2, K1, K2) in the figure matches Eq. (5).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the physical-layer claims are empirical BER benchmarks and the perception-tier gap is an underspecified pipeline, not a self-referential derivation.

full rationale

The paper's derivation chain is empirical rather than analytic. The physical-layer claims (Figs. 2-4) are measured BER comparisons against DeepRx and a conventional baseline under 3GPP UMa and CDL channels, with DeepRx explicitly re-trained under identical conditions; no parameter is fitted to the BER outcome and then renamed a prediction. The perception-tier claim (Table I) inherits BER-quality differences through an underspecified sensor-to-bit encoding and reconstruction pipeline; this is a serious methodological and reproducibility gap that makes the AP values hard to audit, but it is not circular: the AP is not defined in terms of the Diff-Transformer output, and no equation in Section II reduces the perception result to the receiver's training objective. The self-citations [20] and [22] are used only to motivate transformer-based receivers and are not load-bearing, since the performance comparison does not rest on those papers' results. Training and evaluating on the UMa channel is standard same-distribution evaluation rather than a self-definitional loop. Therefore no step in the paper equals its own input by construction, and the appropriate finding is no significant circularity.

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

No new particles, forces, or entities are introduced; the differential attention module is prior art from [19]. The free parameters are architectural hyperparameters chosen without ablation, and the load-bearing assumptions are the simulator fidelity, the HEAL pretrained models, and the unstated sensor-data encoding pipeline.

free parameters (3)

Number of Diff-Transformer blocks = 4
Set by hand without ablation study; the paper does not justify this depth.
Number of attention heads = 4
Chosen by hand; no ablation or sensitivity analysis is provided.
Feedforward network dimension = 128
Chosen by hand; no ablation is reported.

assumptions (4)

domain assumption The 3GPP UMa channel model as implemented in Sionna accurately represents urban V2X propagation.
The channel model is used for both training and testing (Section III-A), so the results depend on the simulator's fidelity.
domain assumption The HEAL framework and its pretrained perception models (OPV2V, OPV2V-H) are valid and applied correctly.
The collaborative perception evaluation relies entirely on HEAL [23] and its pretrained agents (Section III-B) without modification.
ad hoc to paper Perception data (LiDAR and camera) can be serialized into the transmitted bitstream without materially affecting perception quality.
The paper never describes the encoding of sensor data into bits, yet Table I assumes reconstruction preserves enough information for AP evaluation; this assumption is load-bearing.
domain assumption Differential attention from [19] functions as described when applied to raw OFDM samples.
The receiver relies on the published differential attention mechanism (Eq. 5) without further analysis of its behavior on wireless signals.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Differential Transformer-driven 6G Physical Layer for Collaborative Perception Enhancement." pith.science (2026). https://pith.science/paper/BHFAUC6B

@misc{pith2026250620597,
  author       = {Pith},
  title        = {Pith review of: Differential Transformer-driven 6G Physical Layer for Collaborative Perception Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BHFAUC6B}},
  note         = {Machine review of arXiv:2506.20597}
}

read the original abstract

The emergence of 6G wireless networks promises to revolutionize vehicular communications by enabling ultra-reliable, low-latency, and high-capacity data exchange. In this context, collaborative perception techniques, where multiple vehicles or infrastructure nodes cooperate to jointly receive and decode transmitted signals, aim to enhance reliability and spectral efficiency for Connected Autonomous Vehicle (CAV) applications. In this paper, we propose an end-to-end wireless neural receiver based on a Differential Transformer architecture, tailored for 6G V2X communication with a specific focus on enabling collaborative perception among connected autonomous vehicles. Our model integrates key components of the 6G physical layer, designed to boost performance in dynamic and challenging autonomous driving environments. We validate the proposed system across a range of scenarios, including 3GPP-defined Urban Macro (UMa) channel. To assess the model's real-world applicability, we evaluate its robustness within a V2X framework. In a collaborative perception scenario, our system processes heterogeneous LiDAR and camera data from four connected vehicles in dynamic cooperative vehicular networks. The results show significant improvements over state-of-the-art methods, achieving an average precision of 0.84, highlighting the potential of our proposed approach to enable robust, intelligent, and adaptive wireless cooperation for next-generation connected autonomous vehicles.

Figures

Figures reproduced from arXiv: 2506.20597 by the authors.

**Figure 2.** BER vs. SNR performance over a UMa channel with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** BER vs. SNR performance over a UMa channel with [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: BER vs. SNR performance over a CDL channel with [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Visualization Results for Object Detection with 4 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 12 canonical work pages

[1]

Matthaiou, O

M. Matthaiou, O. Yurduseven, H. Q. Ngo, D. Morales-Jimenez, S. L. Cotton & V . F. Fusco, ”The Road to 6G: Ten Physical Layer Challenges for Communications Engineers,” in IEEE Commu- nications Magazine , vol. 59, no. 1, pp. 64-69, January 2021, doi: 10.1109/MCOM.001.2000208

work page doi:10.1109/mcom.001.2000208 2021
[2]

M. K. Shehzad, L. Rose, M. M. Butt, I. Z. Kovacs, M. Assaad & M. Guizani, ”Artificial Intelligence for 6G Networks: Technology Advance- ment and Standardization,” in IEEE V ehicular Technology Magazine, vol. 17, no. 3, pp. 16-25, Sept. 2022, doi: 10.1109/MVT.2022.3164758

arXiv 2022
[3]

& Hadid, A

Ribouh, S. & Hadid, A. SEECAD: Semantic end-to-end communication for autonomous driving. 2024 IEEE Intelligent V ehicles Symposium (IV). pp. 1808-1813 (2024)

work page 2024
[4]

Study on Artificial Intelligence (AI)/Machine Learn- ing (ML) for NR Air Interface RAN,

3GPP Release 18, “Study on Artificial Intelligence (AI)/Machine Learn- ing (ML) for NR Air Interface RAN,” Meeting #112, Athens, Greece, Tech. Rep., 27th February – 3rd March 2023

work page 2023
[5]

& Tatkeu, C

Kushardianto, N., Ribouh, S., El Hillali, Y . & Tatkeu, C. Vehicular network anomaly detection based on 2-step deep learning framework. V ehicular Communications. 49 pp. 100802 (2024)

work page 2024
[6]

& Hadid, A

Ribouh, S. & Hadid, A. Is Semantic Communication for Autonomous Driving Secured against Adversarial Attacks?. 2024 IEEE 6th Interna- tional Conference On AI Circuits And Systems (AICAS) . pp. 139-143 (2024)

work page 2024
[7]

Gharsallah & G

G. Gharsallah & G. Kaddoum, ”MVX-ViT: Multimodal Collaborative Perception for 6G V2X Network Management Decisions Using Vision Transformer,” in IEEE Open Journal of the Communications Society , vol. 5, pp. 5619-5634, 2024, doi: 10.1109/OJCOMS.2024.3452591

arXiv 2024
[8]

Learning the MMSE channel estimator,

D. Neumann, T. Wiese, & W. Utschick, “Learning the MMSE channel estimator,” IEEE Transactions on Signal Processing , vol. 66, no. 11, pp. 2905–2917, June 2018

2018

Show all 25 references

[9]

Deep learning-based channel estima- tion for beamspace mmWave massive MIMO systems,

H. He, C. Wen, S. Jin, & G. Li, “Deep learning-based channel estima- tion for beamspace mmWave massive MIMO systems,” IEEE Wireless Communications Letters , vol. 7, no. 5, 2018

2018
[10]

Complex CNN-based equal- ization for communication signal,

Z. Chang, Y . Wang, H. Li, & Z. Wang, “Complex CNN-based equal- ization for communication signal,” in 2019 IEEE 4th International Conference on Signal and Image Processing (ICSIP) , July 2019, pp. 513–517

2019
[11]

Machine LLRning: Learning to softly demodulate,

O. Shental and J. Hoydis, “Machine LLRning: Learning to softly demodulate,” 2019

2019
[12]

ComNet: Combination of deep learning and expert knowledge in OFDM receivers,

X. Gao, S. Jin, C. Wen, & G. Y . Li, “ComNet: Combination of deep learning and expert knowledge in OFDM receivers,” IEEE Communica- tions Letters , vol. 22, no. 12, pp. 2627–2630, Dec 2018

2018
[13]

Model-driven deep learning for physical layer communications,

H. He, S. Jin, C. Wen, F. Gao, G. Y . Li, & Z. Xu, “Model-driven deep learning for physical layer communications,” IEEE Wireless Communi- cations, vol. 26, no. 5, pp. 77–83, 2019

2019
[14]

Learning to detect,

N. Samuel, T. Diskin, & A. Wiesel, “Learning to detect,” IEEE Transac- tions on Signal Processing , vol. 67, no. 10, pp. 2554–2564, May 2019

2019
[15]

Power of deep learning for channel estimation and signal detection in OFDM systems,

H. Ye, G. Y . Li, & B.-H. Juang, “Power of deep learning for channel estimation and signal detection in OFDM systems,” IEEE Communica- tions Letters , vol. 7, no. 1, pp. 114–117, Feb. 2018

2018
[16]

Z. Zhao, M. C. Vuran, F. Guo & S. D. Scott, ”Deep-Waveform: A Learned OFDM Receiver Based on Deep Complex-Valued Con- volutional Networks,” in IEEE Journal on Selected Areas in Com- munications, vol. 39, no. 8, pp. 2407-2420, Aug. 2021, doi: 10.1109/JSAC.2021.3087241

2021
[17]

Honkala, D

M. Honkala, D. Korpi & 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
[18]

Akrout, A

M. Akrout, A. Feriani, F. Bellili, A. Mezghani & E. Hossain, ”Domain Generalization in Machine Learning Models for Wireless Commu- nications: Concepts, STATE of-the-Art, and Open Issues,” in IEEE Communications Surveys and Tutorials , vol. 25, no. 4, pp. 3014-3037, Fourthquart...

2023
[19]

T. Ye, L. Dong, Y . Xia, Y . Sun, Y . Zhu, G. Huang & F. Wei, 2024. Differential transformer. arXiv preprint arXiv:2410.05258

2024 arXiv
[20]

Saleem, M

O. Saleem, M. Alfaqawi, P. Merdrignac, A. Bensrhair & S. Ribouh, (2025). Deep Multi-modal Neural Receiver for 6G Vehicular Commu- nication. arXiv preprint arXiv:2501.13464

2025 arXiv
[21]

Wiesmayr, S

R. Wiesmayr, S. Cammerer, F. A. Aoudia, J. Hoydis, J. Zakrzewski & A. Keller (2024). Design of a standard-compliant real-time neural receiver for 5G NR. arXiv preprint arXiv:2409.02912

2024 arXiv
[22]

Saleem, S

O. Saleem, S. Ribouh, M. Alfaqawi, A. Bensrhair, & P. Merdrignac, (2024, October). TransRx-6G-V2X: Transformer Encoder-Based Deep Neural Receiver For Next Generation of Cellular Vehicular Communica- tions. In 2024 IEEE 100th V ehicular Technology Conference (VTC2024- Fall) (pp...

2024
[23]

Y . Lu, Y . Hu, Y . Zhong, D. Wang, Y . Wang, & S. Chen (2024). An extensible framework for open heterogeneous collaborative perception. arXiv preprint arXiv:2401.13964

2024 arXiv
[24]

Sionna: An Open-Source Library for Next-Generation Physical Layer Research,

J. Hoydis, S. Cammerer, F. Ait Aoudia, A. Vem, N. Binder, G. Marcus, & A. Keller, “Sionna: An Open-Source Library for Next-Generation Physical Layer Research,” arxiv preprint arXiv:2203.11854, 2022

2022 arXiv
[25]

https://www.atis.org/wp-content/uploads/3gpp- documents/Rel15/ATIS.3GPP.37.885.V1530.pdf

Pith tools

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

[1] [1]

Matthaiou, O

M. Matthaiou, O. Yurduseven, H. Q. Ngo, D. Morales-Jimenez, S. L. Cotton & V . F. Fusco, ”The Road to 6G: Ten Physical Layer Challenges for Communications Engineers,” in IEEE Commu- nications Magazine , vol. 59, no. 1, pp. 64-69, January 2021, doi: 10.1109/MCOM.001.2000208

work page doi:10.1109/mcom.001.2000208 2021

[2] [2]

M. K. Shehzad, L. Rose, M. M. Butt, I. Z. Kovacs, M. Assaad & M. Guizani, ”Artificial Intelligence for 6G Networks: Technology Advance- ment and Standardization,” in IEEE V ehicular Technology Magazine, vol. 17, no. 3, pp. 16-25, Sept. 2022, doi: 10.1109/MVT.2022.3164758

arXiv 2022

[3] [3]

& Hadid, A

Ribouh, S. & Hadid, A. SEECAD: Semantic end-to-end communication for autonomous driving. 2024 IEEE Intelligent V ehicles Symposium (IV). pp. 1808-1813 (2024)

work page 2024

[4] [4]

Study on Artificial Intelligence (AI)/Machine Learn- ing (ML) for NR Air Interface RAN,

3GPP Release 18, “Study on Artificial Intelligence (AI)/Machine Learn- ing (ML) for NR Air Interface RAN,” Meeting #112, Athens, Greece, Tech. Rep., 27th February – 3rd March 2023

work page 2023

[5] [5]

& Tatkeu, C

Kushardianto, N., Ribouh, S., El Hillali, Y . & Tatkeu, C. Vehicular network anomaly detection based on 2-step deep learning framework. V ehicular Communications. 49 pp. 100802 (2024)

work page 2024

[6] [6]

& Hadid, A

Ribouh, S. & Hadid, A. Is Semantic Communication for Autonomous Driving Secured against Adversarial Attacks?. 2024 IEEE 6th Interna- tional Conference On AI Circuits And Systems (AICAS) . pp. 139-143 (2024)

work page 2024

[7] [7]

Gharsallah & G

G. Gharsallah & G. Kaddoum, ”MVX-ViT: Multimodal Collaborative Perception for 6G V2X Network Management Decisions Using Vision Transformer,” in IEEE Open Journal of the Communications Society , vol. 5, pp. 5619-5634, 2024, doi: 10.1109/OJCOMS.2024.3452591

arXiv 2024

[8] [8]

Learning the MMSE channel estimator,

D. Neumann, T. Wiese, & W. Utschick, “Learning the MMSE channel estimator,” IEEE Transactions on Signal Processing , vol. 66, no. 11, pp. 2905–2917, June 2018

2018