X-REFINE: XAI-based RElevance input-Filtering and archItecture fiNe-tuning for channel Estimation

Abdul Karim Gizzini; Yahia Medjahdi

REVIEW 5 major objections 4 minor 12 references

X-REFINE: XAI-based RElevance input-Filtering and archItecture fiNe-tuning for channel Estimation

T0 review · 5 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read X-REFINE claims that a single sign-stabilized LRP relevance map can drive joint pruning of both OFDM subcarriers and hidden neurons in a channel-estimation FNN, cutting computational cost by 35–62% while holding bit-error rate, and that thi

desk verdict X-REFINE is a plausible incremental extension of the authors' earlier XAI-CHEST framework, but the current evidence does not show that LRP-based relevance is what drives the pruning gains. read the letter →

arxiv 2602.22277 v2 pith:7IQBDNM2 submitted 2026-02-25 cs.LG eess.SP

Abdul Karim Gizzini , Yahia Medjahdi This is my paper

classification cs.LGeess.SP

keywords layer-wiserelevancepropagationchannelestimationOFDMinputfilteringarchitecturepruningexplainableAIV2Xcommunicationsneuralnetworkcomplexityreduction

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

The paper tries to show that one explainability signal—layer-wise relevance propagation with sign stabilization—can drive two pruning decisions at once: which OFDM subcarriers feed the neural denoiser and which hidden neurons are removed. It argues that perturbation-based XAI, which only filters inputs, misses structural savings, and that LRP gives sharper relevance maps because it backpropagates through the network. If correct, a wireless receiver can run a much smaller FNN at comparable bit-error rate, which matters for latency- and power-constrained 6G vehicle-to-everything links. The evidence is BER simulations in four scenarios plus a FLOPs count showing 35–62% complexity reduction relative to the full model, with the best results on low-frequency-selective channels.

What carries the argument

The sign-stabilized LRP-epsilon rule (Eq. 6) redistributes relevance backward through the network, splitting each neuron's relevance to its predecessors in proportion to z_ij = a_i w_ij, normalized by successor sums with an epsilon sign-stabilizer. The conservation law (Eq. 5) states that total relevance at any layer equals the norm of the complex output, and this single rule produces both the input relevance vector R_bar_in and per-layer neuron relevance R_bar_arch. These become binary masks via value-based thresholding (inputs) and percentile-based pruning (architecture). The cost model (Eq. 7) counts active connections using l0-norms, so reducing kept subcarriers and neurons translates di

What would settle it

Train the same STA-FNN on synthetic channels where the truly important subcarriers and the truly redundant hidden neurons are known by construction. If the LRP-derived masks do not match those known important and redundant subsets—or if the X-REFINE pruned model's BER degrades sharply when the supposedly irrelevant subcarriers are removed—the central trade-off claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the sign-stabilized LRP-epsilon rule, with relevance initialized as the norm of the complex channel estimate, produces reliable importance scores for both the FNN's 104 input subcarriers and its hidden neurons. Thresholding these scores yields an input mask (keeping subcarriers with relevance above tau) and an architecture mask (pruning the bottom P percent of neurons per layer), and a joint grid search over (tau, P) finds masks with lower FLOPs and BER no worse than the full model. Across LF and HF V2V channels and QPSK/64QAM, the pruned model keeps BER robust, with complexity reductions of 35.16–62.41% versus the full model, consistently beating the perturbation-b

Load-bearing premise

The load-bearing premise is that the layer-wise relevance propagation conservation law and redistribution rule (Eqs. 5–6), originally designed for real-valued networks, remain faithful when applied to a complex-valued FNN with relevance initialized as the norm of the complex output; the paper does not prove or empirically validate this faithfulness.

Editorial extensions

If this is right

Keeping only the pilot subcarriers is sufficient in low-frequency-selective channels, allowing input size to drop from 104 to 8 subcarriers with no BER loss.
Architectural pruning of 15–25% of neurons, concentrated in the middle hidden layer, preserves BER while adding substantial FLOPs savings on top of input filtering.
As modulation order rises from QPSK to 64QAM, the number of zero-relevance subcarriers grows, so the model needs fewer inputs but more network capacity; the optimal pruning percentile falls accordingly.
The framework's complexity metric (Eq. 7) gives a closed-form FLOPs count for any mask pair, so the computation cost of a mask can be evaluated without running the network.

Reading between the lines

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

If the LRP relevance scores are faithful, the same double-mask recipe should transfer to other structured neural-network inputs—MIMO antenna grids, time-domain taps, or frequency-time resources—because the rule itself does not depend on OFDM specifics.
The grid search over (tau, P) is feasible only because the FNN is small (three hidden layers); scaling to wider or deeper networks would likely require a differentiable or learned mask rather than exhaustive search.
The paper treats negative-relevance subcarriers as 'harmful', implying a testable prediction: deliberately re-adding the most negative-relevance subcarriers to the pruned model should measurably degrade BER, which would confirm that the signs carry causal meaning.
If unfaithful relevance were the issue, a natural diagnostic is to compare LRP-derived masks against exhaustive input-ablation and neuron-ablation results on the same trained model; the paper does not provide this validation.

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 proposes X-REFINE, which uses LRP relevance scores to prune both the inputs and hidden neurons of a channel-estimation FNN. That is a real step beyond the authors' earlier perturbation-based XAI-CHEST, and the complexity accounting is clear and reproducible. The reported FLOPs reductions of 35–62% are concrete, and the experiments cover two modulations and two V2V channel models. The writing is mostly clear, and the authors are honest that the joint optimization is a grid search, even though the contribution list overstates it as an analytical expression.

The load-bearing weakness is the LRP extension. Equation (5) asserts conservation with respect to the norm of the complex output, and Equation (6) applies the epsilon rule without specifying how the sign and division are defined for complex quantities. If the FNN actually operates on real and imaginary components separately—which the input mask dimension suggests—that needs to be stated explicitly, and the norm initialization must be justified. More importantly, the faithfulness of the resulting relevance scores is never validated. That pilots receive high relevance is unsurprising and does not confirm that the ranking of other subcarriers or hidden neurons is meaningful. Without that validation, the masks could be arbitrary.

Second, Algorithm 1 selects masks based on validation loss and a BER target, so the claim that pruned models maintain BER is partly guaranteed by the selection criterion. That is acceptable if the paper shows that LRP-guided masks outperform random or magnitude-based pruning at the same compression level, but no such baseline is included. Without it, the unique contribution of LRP is untested. Third, there are no error bars or multiple runs, and the BER values are not stated numerically in the text.

These are addressable issues. The paper is coherent on its own terms, the citation pattern is fine, and the core idea is worth exploring. I would send it to peer review with the expectation of major revisions: validate the LRP scores against ground-truth importance, add simple pruning baselines, and report variance. It is not a groundbreaking result, but for researchers applying XAI to wireless model compression it is a legitimate contribution once the evidence is tightened.

Referee Report

5 major / 4 minor

Summary. The paper proposes X-REFINE, a framework that uses a sign-stabilized layer-wise relevance propagation (LRP-epsilon) rule to jointly prune the input features and hidden units of a feed-forward neural network used for OFDM channel estimation. The network is first trained on noisy initial channel estimates; LRP relevance scores are then computed for subcarriers and for neurons in each hidden layer, and binary masks are derived by thresholding the input scores and by percentile-based pruning of the hidden neurons. The masked model is retrained, and the best mask pair is selected by a grid search over the threshold and pruning percentile, subject to validation-loss improvement and a BER target. Simulations on IEEE 802.11p V2V scenarios (LF/HF channels, QPSK/64QAM) show FLOPs reductions of 35-62% relative to the full model and improved BER compared with the perturbation-based XAI-CHEST baseline.

Significance. If the central claim holds, X-REFINE would be a useful contribution: it provides a concrete, model-agnostic recipe for using decomposition-based XAI not only to filter inputs but also to prune internal architecture, with explicit FLOPs accounting and evaluation across several wireless scenarios. The idea of applying LRP to complex-valued channel-estimation networks and using the resulting relevance to guide joint pruning is timely for 6G AI-native research. The manuscript also has strengths: it states an explicit optimization objective, gives a precise FLOPs model, and compares against an existing XAI method. However, the significance is currently conditional: the LRP extension to complex-valued regression is not rigorously defined or validated, the selection procedure in Algorithm 1 partly guarantees the reported BER preservation, and the lack of random/magnitude-pruning baselines means the specific contribution of LRP is untested.

major comments (5)

[§III-B, Eqs. (5)-(6)] The LRP rule is transplanted to a complex-valued FNN without the necessary definitions. In Eq. (6), z_ij, the denominator sum_k z_kj, and sign() are complex quantities; the manuscript does not define complex division, a complex sign function, or how the complex-valued relevance R_i is reduced to a scalar for thresholding. The norm-initialization of relevance in Eq. (5) is also nonstandard and is not justified beyond a heuristic statement. Since the masks m_in and m_arch are derived directly from these scores, the central claim collapses if the scores are not faithful. The only empirical validation offered (pilots receive high relevance) is circular because pilots are known a priori to be important. I request either (a) a rigorous statement of the complex-valued LRP rule with all operations defined, or (b) a validation of the relevance scores against a ground-truth attribution, e.g., by c
[§III-A, Algorithm 1] Algorithm 1 selects m* only when Val-Loss_c < Val-Loss_p and BER ≤ BER_target (lines 8-10). Thus the claim that pruned models 'maintain robust BER performance' is partly guaranteed by the selection criterion rather than being an independent discovery. The reported BER curves are for the selected mask only; there is no distribution over initializations or retraining runs. To make the claim load-bearing, the authors should report the average and spread of BER over multiple random seeds and, ideally, show the BER of a fixed-compression mask chosen without the validation-based selection loop.
[Introduction, Contribution 2 vs. §III-A] The Introduction states as a contribution 'Deriving the analytical expression and the corresponding simulations of both the model inputs and architecture optimization.' However, §III-A says the problem is transformed into a structured numerical grid search and explicitly states that no closed-form derivation is required. Equation (4) is an optimization problem, not an analytical solution. This is internally inconsistent and should be corrected: either present the grid search as the method, or derive an actual closed-form threshold/per percentile expression.
[§IV-B and §IV-C] The BER claims are not numerically stated in the text and no error bars or multiple runs are reported. For example, 'slightly improving the BER performance' (QPSK, LF) and 'preserving the best BER performance' (P=25%) are qualitative statements. With a single run and no confidence intervals, the reader cannot judge whether the observed differences are within noise. Tables with mean BER and standard deviation, or at least error bars on the BER curves, are needed to support the performance claims.
[§IV, overall evaluation] There is no comparison to random pruning or magnitude-based pruning at the same compression ratio. X-REFINE is compared only against XAI-CHEST and the full baseline. Since a random mask at the same sparsity could achieve similar BER after retraining, the claimed superiority of LRP-based relevance is not established. I request a control experiment: for each scenario, prune the same number of inputs and neurons using (a) random masks and (b) weight-magnitude or activation-based saliency, and report the resulting BER. Without this, the unique contribution of the LRP mechanism remains untested.

minor comments (4)

[§III-A, Eq. (3)-(4)] There is a dimensional inconsistency: Eq. (3) uses m_in with the same dimension as the complex channel vector in C^{K_on}, while the text sets m_in ∈ {0,1}^{2K_on}. If the input is split into real and imaginary parts, this should be stated explicitly; otherwise the mask dimension is wrong.
[§II] Typo: 'e_q demotes the Doppler-induced inter-carrier interference' should be 'denotes'.
[§IV-D, Figure 4] The FLOPs percentages in Figure 4 are not explicitly tied to the masks reported in §IV-B/C. For reproducibility, list the exact input mask and per-layer architectural masks (e.g., (14-11-09)) in a table for each scenario, along with the corresponding FLOPs from Eq. (7).
[§III-B] The phrase 'holistic optimization' in the Abstract is not used consistently with the 'double optimization' terminology in the paper; consider aligning the wording.

Circularity Check

1 steps flagged · score 4.0 of 10

BER maintenance is partly enforced by Algorithm 1's acceptance rule; the LRP-to-mask derivation is otherwise not circular.

fitted input called prediction [Algorithm 1 / Section III-A]
"if Val-Loss c < Val-Loss p and BER≤BER target then m∗ in ←m in, m∗ arch ←m arch"

The BER target is an input to the mask-acceptance rule, so the reported BER of the optimized model is not an independent discovery of the LRP-based pruning; any mask failing the BER target is discarded. Thus the conclusion 'maintaining robust BER while reducing complexity' is partly guaranteed by the selection criterion rather than demonstrated by the XAI mechanism. The complexity reduction itself is arithmetic from the accepted masks. The independent content is limited to showing that some LRP-derived masks satisfied the acceptance criterion, which is not compared against random or magnitude-based pruning.

full rationale

The paper's core derivation—LRP relevance (Eqs. 5–6) to masks (Eq. 4) to FLOP reduction (Eq. 7)—is not formally circular: the LRP rule is imported from an external reference [9], and the masks are explicit functions of relevance scores. The main circularity concern is Algorithm 1, where a candidate mask is retained only when validation loss improves and BER ≤ BER target. Therefore the 'robust BER' result is partly a selection outcome rather than an independent prediction of the LRP mechanism. The self-citations to XAI-CHEST [8] and to [10], [11] are used as baselines and system components; they are not load-bearing for the central derivation and do not by themselves create circularity. The unvalidated extension of real-valued LRP to complex-valued regression is a correctness risk (the relevance scores may be unfaithful), but the paper states the initialization explicitly, so it is not a hidden circular step. No uniqueness theorem or hidden ansatz is imported from the authors' prior work. Because the cross-scenario comparisons and complexity analysis retain independent content, the score is 4 rather than higher.

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

The central mechanism rests on the validity of LRP for complex-valued networks, the channel model, and the choice of thresholds; none of these is independently established in this paper beyond citation and validation-set selection.

free parameters (3)

Input relevance threshold τ = grid search over T; optimal values not reported
Controls which subcarriers are kept; tuned per scenario/modulation on validation loss (Eq. (4), Algorithm 1).
Architecture pruning percentile P = optimal P in {15,20,25,30}, e.g., 25% (LF-QPSK), 15% (other tested cases)
Selected by grid search on validation; directly determines the reported complexity reductions.
LRP epsilon stabilizer = unspecified
Introduced ad hoc in Eq. (6) as a sign-dependent stabilizer; no value or sensitivity analysis given, and it affects the relevance distribution.

assumptions (4)

domain assumption OFDM signal model with doubly-selective channel and Doppler-induced ICI (Eq. (1) and [10]).
The evaluation is built on this prior channel/ICI model; errors in that model propagate into all simulation results.
ad hoc to paper LRP conservation law (Eq. (5)) and redistribution formula (Eq. (6)) hold for complex-valued FNN with norm-initialized relevance.
Standard LRP theory is for real-valued networks; this paper-specific extension is unproven and load-bearing.
domain assumption Percentile-based pruning preserves network integrity; relevance scores scale down with depth.
Asserted in Section III-A without proof; if scores do not scale uniformly, fixed percentile pruning could remove the wrong neurons.
domain assumption Value-based thresholding exploits channel sparsity; subcarriers below τ are safe to remove.
This sparsity assumption justifies input filtering but is not independently verified in the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of X-REFINE: XAI-based RElevance input-Filtering and archItecture fiNe-tuning for channel Estimation." pith.science (2026). https://pith.science/paper/7IQBDNM2

@misc{pith2026260222277,
  author       = {Pith},
  title        = {Pith review of: X-REFINE: XAI-based RElevance input-Filtering and archItecture fiNe-tuning for channel Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7IQBDNM2}},
  note         = {Machine review of arXiv:2602.22277}
}

read the original abstract

AI-native architectures are vital for 6G wireless communications. The black-box nature and high complexity of deep learning models employed in critical applications, such as channel estimation, limit their practical deployment. While perturbation-based eXplainable Artificial Intelligence (XAI) solutions offer input filtering, they often neglect internal structural optimization. We propose X-REFINE, an XAI-based framework for joint input-filtering and architecture fine-tuning. By utilizing a decomposition-based, sign-stabilized LRP epsilon rule, X-REFINE backpropagates predictions to derive high-resolution relevance scores for both subcarriers and hidden neurons. This enables a reliable optimization that identifies the most reliable model components. Simulation results demonstrate that X-REFINE achieves a superior performance-complexity-interpretability trade-off compared to the external perturbation-based XAI frameworks, significantly reducing computational complexity while maintaining robust bit error rate (BER) performance.

Figures

Figures reproduced from arXiv: 2602.22277 by the authors.

**Figure 2.** Relevance Scores distribution of XAI-CHEST, and the proposed X-REFINE framework considering the LF and HF channel models under different [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** BER performance for the LF and HF channel models under different modulation schemes. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Complexity comparison across different scenarios. The baseline is [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 1 linked inside Pith

[1]

Toward Smart 5G and 6G: Standardization of AI-Native Network Architectures and Semantic Communication Protocols,

K. Jain, P. Krishnan, P. Pachiyannan, L. Jaganathan, M. A. Khan, and Y . Li, “Toward Smart 5G and 6G: Standardization of AI-Native Network Architectures and Semantic Communication Protocols,”IEEE Communications Standards Magazine, pp. 1–12, 2025

2025
[2]

6G-Enabled Vehicle-to-Everything Communications: Current Research Trends and Open Challenges,

J. Rodr ´ıguez-Pi˜neiro, Z. Wei, J. Wang, C. A. Guti ´errez, and L. M. Correia, “6G-Enabled Vehicle-to-Everything Communications: Current Research Trends and Open Challenges,”IEEE Open Journal of V ehicular Technology, vol. 6, pp. 2358–2391, 2025

2025
[3]

Sovereign AI for 6G: Towards the Future of AI-Native Networks,

S. B. Chetty, D. Grace, S. Saunders, P. Harris, E. E. Tsiropoulou, T. Quek, and H. Ahmadi, “Sovereign AI for 6G: Towards the Future of AI-Native Networks,”arXiv preprint arXiv:2509.06700, 2025

arXiv 2025
[4]

Toward 6G: An Overview of the Next Generation of Intelligent Network Connectivity,

S. Prasad Tera, R. Chinthaginjala, G. Pau, and T. Hoon Kim, “Toward 6G: An Overview of the Next Generation of Intelligent Network Connectivity,”IEEE Access, vol. 13, pp. 925–961, 2025

2025
[5]

On Combining XAI and LLMs for Trustworthy Zero-Touch Network and Service Management in 6G,

A. Mekrache, M. Mekki, A. Ksentini, B. Brik, and C. Verikoukis, “On Combining XAI and LLMs for Trustworthy Zero-Touch Network and Service Management in 6G,”IEEE Communications Magazine, vol. 63, no. 4, pp. 154–160, 2025

2025
[6]

Advancing 6G: Survey for Explainable AI on Communications and Network Slicing,

H. Sun, Y . Liu, A. Al-Tahmeesschi, A. Nag, M. Soleimanpour, B. Can- berk, H. Arslan, and H. Ahmadi, “Advancing 6G: Survey for Explainable AI on Communications and Network Slicing,”IEEE Open Journal of the Communications Society, vol. 6, pp. 1372–1412, 2025

2025
[7]

A Survey on XAI for 5G and Beyond Security: Technical Aspects, Challenges and Research Directions,

T. Senevirathna, V . H. La, S. Marcha, B. Siniarski, M. Liyanage, and S. Wang, “A Survey on XAI for 5G and Beyond Security: Technical Aspects, Challenges and Research Directions,”IEEE Communications Surveys & Tutorials, vol. 27, no. 2, pp. 941–973, 2025

2025
[8]

Explainable AI for Enhancing Efficiency of DL-based Channel Estimation,

A. K. Gizzini, Y . Medjahdi, A. J. Ghandour, and L. Clavier, “Explainable AI for Enhancing Efficiency of DL-based Channel Estimation,”IEEE Transactions on Machine Learning in Communications and Networking, 2025

2025

Show all 12 references

[9]

Layer-Wise Relevance Propagation: An Overview,

G. Montavon, A. Binder, S. Lapuschkin, W. Samek, and K.-R. M ¨uller, “Layer-Wise Relevance Propagation: An Overview,”Explainable AI: interpreting, explaining and visualizing deep learning, pp. 193–209, 2019

2019
[10]

Low Complex Methods for Robust Channel Estimation in Doubly Dispersive Environments,

A. K. Gizzini and M. Chafii, “Low Complex Methods for Robust Channel Estimation in Doubly Dispersive Environments,”IEEE Access, vol. 10, pp. 34 321–34 339, 2022

2022
[11]

A Survey on Deep Learning Based Channel Estimation in Doubly Dispersive Environments,

——, “A Survey on Deep Learning Based Channel Estimation in Doubly Dispersive Environments,”IEEE Access, vol. 10, pp. 70 595–70 619, 2022

2022
[12]

Vehicle–Vehicle Channel Models for the 5- GHz Band,

I. Sen and D. W. Matolak, “Vehicle–Vehicle Channel Models for the 5- GHz Band,”IEEE Transactions on Intelligent Transportation Systems, vol. 9, no. 2, pp. 235–245, 2008

2008

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Toward Smart 5G and 6G: Standardization of AI-Native Network Architectures and Semantic Communication Protocols,

K. Jain, P. Krishnan, P. Pachiyannan, L. Jaganathan, M. A. Khan, and Y . Li, “Toward Smart 5G and 6G: Standardization of AI-Native Network Architectures and Semantic Communication Protocols,”IEEE Communications Standards Magazine, pp. 1–12, 2025

2025

[2] [2]

6G-Enabled Vehicle-to-Everything Communications: Current Research Trends and Open Challenges,

J. Rodr ´ıguez-Pi˜neiro, Z. Wei, J. Wang, C. A. Guti ´errez, and L. M. Correia, “6G-Enabled Vehicle-to-Everything Communications: Current Research Trends and Open Challenges,”IEEE Open Journal of V ehicular Technology, vol. 6, pp. 2358–2391, 2025

2025

[3] [3]

Sovereign AI for 6G: Towards the Future of AI-Native Networks,

S. B. Chetty, D. Grace, S. Saunders, P. Harris, E. E. Tsiropoulou, T. Quek, and H. Ahmadi, “Sovereign AI for 6G: Towards the Future of AI-Native Networks,”arXiv preprint arXiv:2509.06700, 2025

arXiv 2025

[4] [4]

Toward 6G: An Overview of the Next Generation of Intelligent Network Connectivity,

S. Prasad Tera, R. Chinthaginjala, G. Pau, and T. Hoon Kim, “Toward 6G: An Overview of the Next Generation of Intelligent Network Connectivity,”IEEE Access, vol. 13, pp. 925–961, 2025

2025

[5] [5]

On Combining XAI and LLMs for Trustworthy Zero-Touch Network and Service Management in 6G,

A. Mekrache, M. Mekki, A. Ksentini, B. Brik, and C. Verikoukis, “On Combining XAI and LLMs for Trustworthy Zero-Touch Network and Service Management in 6G,”IEEE Communications Magazine, vol. 63, no. 4, pp. 154–160, 2025

2025

[6] [6]

Advancing 6G: Survey for Explainable AI on Communications and Network Slicing,

H. Sun, Y . Liu, A. Al-Tahmeesschi, A. Nag, M. Soleimanpour, B. Can- berk, H. Arslan, and H. Ahmadi, “Advancing 6G: Survey for Explainable AI on Communications and Network Slicing,”IEEE Open Journal of the Communications Society, vol. 6, pp. 1372–1412, 2025

2025

[7] [7]

A Survey on XAI for 5G and Beyond Security: Technical Aspects, Challenges and Research Directions,

T. Senevirathna, V . H. La, S. Marcha, B. Siniarski, M. Liyanage, and S. Wang, “A Survey on XAI for 5G and Beyond Security: Technical Aspects, Challenges and Research Directions,”IEEE Communications Surveys & Tutorials, vol. 27, no. 2, pp. 941–973, 2025

2025

[8] [8]

Explainable AI for Enhancing Efficiency of DL-based Channel Estimation,

A. K. Gizzini, Y . Medjahdi, A. J. Ghandour, and L. Clavier, “Explainable AI for Enhancing Efficiency of DL-based Channel Estimation,”IEEE Transactions on Machine Learning in Communications and Networking, 2025

2025