REVIEW 4 major objections 4 minor 15 references
Feature Engineering for Wireless Communications and Networking: Concepts, Methodologies, and Applications
T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read The paper proposes a diffusion-based feature transformation framework that reconstructs attack-free RSSI feature spectra from jammer-corrupted measurements in low-altitude ISAC networks, reporting SSIM improvements of 8 percent to 87…
desk verdict Solid survey, but the case study's central claim is undermined by a 4% vs 44% discrepancy and an uncharacterized 'submerging' assumption; worth refereeing, needs major revision. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the RSSI feature spectrum, a 128 by 128 spatial map of received signal strength over a low-altitude grid, and the recovery engine is a Markov-chain diffusion model with two processes. The forward process injects Gaussian noise with a tuned step count so that attack perturbations are submerged while the signal pattern is not erased; the reverse process, implemented with an attention-equipped U-Net and time embeddings, denoises the spectrum. Multiple forward-reverse rounds and using the perturbed spectrum as guidance let the framework find a path back to the clean spectrum manifold, with structural similarity evaluation used as the reference in the case study.
What would settle it
Construct a low-altitude scenario with the same settings but use a spoofing attack that adds a strong, spatially localized false peak (not diffuse Gaussian-like jamming) at attack probability 0.3. If the reconstructed spectrum's SSIM against the attack-free baseline does not improve by roughly the reported 8 percent or more, or if a simple low-pass filter of the attacked spectrum matches the diffusion output's SSIM, the claim that diffusion specifically removes attack perturbations would be falsified.
Extended reading notes
Core claim
The central claim is that a generative diffusion model, pretrained only on clean RSSI feature spectra, can act as a feature transformation module that removes both injected Gaussian noise and attack-induced corruption from an attacked spectrum. The forward process adds controlled Gaussian noise to 'submerge' the attack perturbations, and the reverse process, built from a U-Net with attention and time embeddings, denoises iteratively to return to the clean-spectrum manifold. Using the attacked spectrum itself as guidance and running multiple forward-reverse rounds balances submerging perturbations against destroying the underlying signal. The paper reports that reconstruction raises SSIM against the attack-free reference in all tested cases: smallest improvement 8 percent for a ground attack at probability 0.3, largest 87 percent for an airborne attack at probability 0.3, and an average of 44 percent. The abstract states an average improvement of 4 percent, an internal discrepancy with the full-text evaluation.
Load-bearing premise
The whole recovery claim rests on the assumption that attack-induced corruption looks enough like the Gaussian noise the diffusion model was trained to remove that a denoiser trained only on clean spectra will push attacked spectra back to the clean manifold; the paper does not demonstrate this for non-Gaussian or structured attacks.
Editorial extensions
If this is right
- ISAC services that rely on RSSI spectra—localization, trajectory optimization, communication-quality assessment—can keep working while a jammer is active, as long as the attack probability stays below roughly 0.5.
- Because the diffusion model is trained only on clean spectra, the framework does not need examples of attack signatures during training; an operator could deploy it before knowing the adversary's strategy.
- The method's performance envelope is explicit: attack probabilities above 0.5 degrade recovery, and airborne attacks within line of sight are the harder case, so deployment should budget for those conditions.
- Feature transformation can be treated as a generative task: turning a corrupted engineered feature into a clean one is the same operation as denoising, which generalizes the role of feature engineering from preprocessing to defensive reconstruction.
Reading between the lines
- A natural testable extension is to replace the Gaussian-noise forward process with attack-aware perturbations (e.g., localized spoofed peaks) during training; if the paper's premise is right, the framework should still reconstruct clean spectra, and if attacks are structured, denial-of-service-like corruption might survive denoising.
- Editorially, the 44 percent average SSIM figure should be treated as a simulation result under the 3GPP TR 38.901 path-loss model, not a field measurement; real deployment would need to verify recovery against hardware impairments and non-stationary jamming.
- The framework's success suggests a broader principle: any engineered feature that lives on a low-dimensional manifold could be defended by diffusion purification, not just RSSI spectra, so the same method could be tried on CSI tensors or micro-Doppler spectrograms.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This manuscript is a survey of feature engineering for AI-driven wireless communications, with emphasis on low-altitude integrated sensing and communication (ISAC) networks. After reviewing feature creation, transformation, and selection, and their uses in signal processing, resource optimization, anomaly detection, and ISAC, the paper proposes a diffusion-based framework for recovering RSSI feature spectra corrupted by ground-based or airborne jamming. A case study simulates an eVTOL collecting 128×128 RSSI spectra under attack probabilities from 0.3 to 0.7 and reports SSIM improvements after reconstruction. The central claim is that the proposed feature transformation framework can reconstruct signal feature spectra that closely match attack-free baselines, supporting downstream ISAC applications.
Significance. If the reported results are correct, the proposed framework would be a meaningful step toward making low-altitude ISAC feature representations robust to jamming, and the survey portions provide a useful organized overview of feature engineering in wireless systems. The paper also makes a conceptually interesting proposal: using a diffusion forward process to 'submerge' structured attack perturbations and a clean-only trained reverse process to purify them. However, the contribution's validity currently rests on a case study that is incompletely specified and whose headline numbers are internally inconsistent. The manuscript does not provide reproducible training details, baselines, or out-of-distribution evaluation, so the quantitative claim is not yet established.
major comments (4)
- [Abstract vs. Section V.C] The abstract states that the case study achieves an average SSIM improvement of 4%, while Section V.C reports that 'the smallest improvement is 8%, the largest improvement reaches 87%, and the average improvement is 44%.' These numbers cannot both describe the same experiments. Because the paper's central quantitative claim is the reconstruction improvement, this discrepancy must be resolved and a single set of consistent, reproducible numbers reported.
- [Section V.A] The proposed framework is not specified at the level needed to reproduce or evaluate it. The text mentions diffusion steps, noise levels sigma_1, sigma_2, sigma_3 (appearing in Fig. 3), multiple forward-reverse rounds, and SSIM-guided reverse steps, but no training loss, optimizer, hyperparameters, algorithm pseudocode, or guidance-weight formula is given. In particular, the load-bearing premise that adding Gaussian noise 'submerges' a 10 dBm structured jamming signal, and that a reverse process trained only on clean spectra then removes both injected noise and attack corruption, is never characterized. The manuscript should provide the noise schedule, the interference-to-noise ratio relative to the jamming power, and an analysis or experiment showing that attacked spectra after forward noising are statistically indistinguishable from noised clean spectra.
- [Section V.C] The evaluation uses attack-free spectra as the SSIM reference, and the diffusion model is pretrained on clean signal features, so the reconstruction target is the model's own training distribution. This is an in-sample demonstration; there is no out-of-distribution test with, for example, attack geometries, jamming powers, or channel conditions not seen during training. The claim that the framework reconstructs 'the attack-free baseline' for each specific scene requires evidence that the model does not merely output a generic clean-looking spectrum. An ablation or baseline comparison (e.g., Gaussian filtering, median filtering, a denoising autoencoder, or purification without guidance) and a sensitivity analysis over random seeds and attack configurations are also missing.
- [Section V.C, Fig. 5] Figure 5 does not allow the reader to verify the reported percentages. The numerical values listed in the text (0.44, 0.40, 0.36, 0.37, 0.35 for attacked spectra and 0.82, 0.72, 0.63, 0.61, 0.59 for reconstructed spectra) do not transparently yield the stated 8%–87% improvements, and the figure legend contains typographical errors ('Reonstructed'). The authors should provide a table of exact SSIM values for every attack scenario and probability, with error bars over multiple simulation runs.
minor comments (4)
- [Throughout] The manuscript contains numerous typographical errors, including 'airspce' (Section V.B), 'Reonstructed' (Fig. 5), 'Beside' (Section VI.B), and inconsistent spacing such as 'UA V' instead of 'UAV'. A careful proofreading pass is needed.
- [Fig. 3] The mathematical notation in Figure 3 is partially garbled (e.g., expressions involving sigma and x with subscripts and superscripts), making the figure hard to interpret. The schematic should be redrawn with clear, standard diffusion equations or a clean block diagram.
- [Section V.A] The roles of 'feature creation' and 'feature transformation' in the proposed framework are described loosely; for example, noise injection is called feature creation, while the U-Net plus denoising is called a feature transformation module. Clarifying how these map to the definitions in Section II would strengthen the connection to the survey's conceptual framework.
- [General] No code, data, or training details are provided, and the case study's reproducibility is therefore limited. Even a small public repository or a detailed implementation appendix would substantially improve the paper.
Circularity Check
No circularity found: the reconstruction case study is a standard in-domain denoising evaluation, with validation gaps but no by-construction reduction to its inputs.
full rationale
The paper's central claim is that a diffusion model, pretrained on clean RSSI feature spectra, can reconstruct an attack-free spectrum from an attacked one by first adding Gaussian noise to 'submerge' the attack and then reverse-denoising. This is an empirical claim about a generative denoising method, not a derived identity. The evaluation measures SSIM against the attack-free spectrum, which is a standard ground-truth reference; the model is not fitted to that reference, and no fitted parameter is renamed as a prediction. The load-bearing premise—that Gaussian noise injection makes attacked spectra indistinguishable from noised clean spectra so that a clean-trained reverse process can remove attack corruption—is asserted rather than demonstrated, and the paper lacks a noise schedule or interference-to-noise analysis. That is a validation weakness, not circularity. The abstract's 4% average SSIM improvement versus Section V.C's 44% (and the misattributed smallest-improvement case) is an internal consistency error, again not a circularity. Reference [1], a prior guided-diffusion adversarial-purification preprint by an overlapping author, is cited as background for diffusion models but is not invoked as the proof of the framework's effectiveness; the synthetic case study is a separate, albeit limited, demonstration. Because no equation is shown to equal its own input and no parameter fitted to the test target is relabeled as a prediction, no specific circular step can be exhibited under the required standard.
Assumptions & free parameters
free parameters (4)
- number of diffusion steps =
100 (from Fig. 3)
- noise levels sigma_1, sigma_2, sigma_3 =
implied by Fig. 3 notation, values not stated
- number of forward-reverse rounds =
not specified
- guidance weight for attacked features =
not specified
assumptions (4)
- domain assumption 3GPP TR 38.901 channel model with LoS probability coefficients 9.61 and 0.16
- domain assumption Logarithmic path-loss model with LoS exponent 2.2, NLoS exponent 3.8, and shadow fading N(0, 6 dB) with 50 m de-correlation
- ad hoc to paper Attack perturbations can be removed by Gaussian denoising trained on clean spectra
- domain assumption RSSI spectrum over a 128x128 grid is a sufficient feature representation for downstream ISAC
Cite this review
Pith. "Pith review of Feature Engineering for Wireless Communications and Networking: Concepts, Methodologies, and Applications." pith.science (2026). https://pith.science/paper/MNVO5WNE
@misc{pith2026250719837,
author = {Pith},
title = {Pith review of: Feature Engineering for Wireless Communications and Networking: Concepts, Methodologies, and Applications},
year = {2026},
howpublished = {\url{https://pith.science/paper/MNVO5WNE}},
note = {Machine review of arXiv:2507.19837}
}
read the original abstract
AI-enabled wireless communications have attracted tremendous research interest in recent years, particularly with the rise of novel paradigms such as low-altitude integrated sensing and communication (ISAC) networks. Within these systems, feature engineering plays a pivotal role by transforming raw wireless data into structured representations suitable for AI models. Hence, this paper offers a comprehensive investigation of feature engineering techniques in AI-driven wireless communications. Specifically, we begin with a detailed analysis of fundamental principles and methodologies of feature engineering. Next, we present its applications in wireless communication systems, with special emphasis on ISAC networks. Finally, we introduce a generative AI-based framework, which can reconstruct signal feature spectrum under malicious attacks in low-altitude ISAC networks. The case study shows that it can effectively reconstruct the signal spectrum, achieving an average structural similarity index improvement of 4%, thereby supporting downstream sensing and communication applications.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[1]
Guided diffusion model for adversarial purification,
J. Wang, Z. Lyu, D. Lin, B. Dai, and H. Fu, “Guided diffusion model for adversarial purification,” arXiv preprint arXiv:2205.14969 , 2022
arXiv 2022
-
[2]
Y . Wei, M.-M. Zhao, and M.-J. Zhao, “Channel distribution learning: Model-driven gan-based channel modeling for irs-aided wireless com- munication,” IEEE Transactions on Communications, vol. 70, no. 7, pp. 4482–4497, 2022
work page 2022
-
[3]
Improved edge detection algorithm for canny operator,
Y . Li and B. Liu, “Improved edge detection algorithm for canny operator,” in 2022 IEEE 10th Joint International Information Technology and Artificial Intelligence Conference (ITAIC) , vol. 10. IEEE, 2022, pp. 1–5
work page 2022
-
[4]
Acceleration estimation of signal propagation path length changes for wireless sensing,
J. Wang, H. Du, D. Niyato, M. Zhou, J. Kang, and H. V . Poor, “Acceleration estimation of signal propagation path length changes for wireless sensing,” IEEE Transactions on Wireless Communications , 2024
work page 2024
-
[5]
Multi-frequency based csi compression for vehicle localization in intelligent transportation system,
X. Yang, M. Gao, L. Xie, and M. Zhou, “Multi-frequency based csi compression for vehicle localization in intelligent transportation system,” IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 3, pp. 2719–2732, 2023
work page 2023
-
[6]
Multiview pca: A methodology of feature extraction and dimension reduction for high-order data,
Z. Xia, Y . Chen, and C. Xu, “Multiview pca: A methodology of feature extraction and dimension reduction for high-order data,” IEEE Transactions on Cybernetics , vol. 52, no. 10, pp. 11 068–11 080, 2021
work page 2021
-
[7]
L. Sun, T. Yin, W. Ding, Y . Qian, and J. Xu, “Feature selection with missing labels using multilabel fuzzy neighborhood rough sets and maximum relevance minimum redundancy,”IEEE Transactions on Fuzzy Systems, vol. 30, no. 5, pp. 1197–1211, 2021
work page 2021
-
[8]
Joint signal detection and automatic modulation classification via deep learning,
H. Xing, X. Zhang, S. Chang, J. Ren, Z. Zhang, J. Xu, and S. Cui, “Joint signal detection and automatic modulation classification via deep learning,” IEEE Transactions on Wireless Communications , 2024
work page 2024
Show all 15 references
-
[9]
An adaptive beamforming technique based on cyclostationary signal properties,
L. Castedo and A. R. Figueiras-Vidal, “An adaptive beamforming technique based on cyclostationary signal properties,”IEEE Transactions on Signal Processing , vol. 43, no. 7, pp. 1637–1650, 2002
2002
-
[10]
Power control based on deep reinforcement learning for spectrum sharing,
H. Zhang, N. Yang, W. Huangfu, K. Long, and V . C. Leung, “Power control based on deep reinforcement learning for spectrum sharing,” IEEE Transactions on Wireless Communications , vol. 19, no. 6, pp. 4209–4219, 2020
2020
-
[11]
Mlts-adcns: Machine learning techniques for anomaly detection in communication networks,
H. W. Oleiwi, D. N. Mhawi, and H. Al-Raweshidy, “Mlts-adcns: Machine learning techniques for anomaly detection in communication networks,” IEEE Access, vol. 10, pp. 91 006–91 017, 2022
2022
-
[12]
Generative artificial intelligence assisted wireless sensing: Human flow detection in practical communication environments,
J. Wang, H. Du, D. Niyato, Z. Xiong, J. Kang, B. Ai, Z. Han, and D. I. Kim, “Generative artificial intelligence assisted wireless sensing: Human flow detection in practical communication environments,” IEEE Journal on Selected Areas in Communications , 2024
2024
-
[13]
Resource allocation for uav swarm-assisted green isac networks via multi-agent rl,
Q. Zhu, R. Liu, Q. Liu, and C. Chen, “Resource allocation for uav swarm-assisted green isac networks via multi-agent rl,” IEEE Transac- tions on Green Communications and Networking , 2024
2024
-
[14]
Integrated sensing and communi- cations for uav assisted internet of things based on deep reinforcement learning,
X. Liu, J. Wu, C. Zhao, and Z. Liu, “Integrated sensing and communi- cations for uav assisted internet of things based on deep reinforcement learning,” IEEE Transactions on Vehicular Technology, 2025
2025
-
[15]
From ground to sky: Architectures, applications, and challenges shaping low-altitude wireless networks,
W. Yuan, Y . Cui, J. Wang, F. Liu, G. Sun, T. Xiang, J. Xu, S. Jin, D. Niy- ato, S. Coleri et al. , “From ground to sky: Architectures, applications, and challenges shaping low-altitude wireless networks,” arXiv preprint arXiv:2506.12308, 2025
2025 arXiv
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.