arxiv: 2605.01705 · v1 · submitted 2026-05-03 · 💻 cs.NI · cs.CR· cs.LG· eess.SP

Recognition: unknown

Toward Resilient 5G Networks: Comparative Analysis of Federated and Centralized Learning for RF Jamming Detection

Burak Kantarci, Mohammadreza Amini, Samhita Kuili

Pith reviewed 2026-05-09 17:00 UTC · model grok-4.3

classification 💻 cs.NI cs.CRcs.LGeess.SP

keywords federated learningjamming detection5G networksRF signalsIQ samplesconvolutional neural networkprivacy preservationFedAvg

0 comments

The pith

Federated learning on local IQ samples detects RF jamming in 5G networks at 97 percent accuracy while preserving user equipment privacy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes a federated learning framework that trains a jamming detector across user devices by exchanging only model updates rather than raw radio signals. It extracts In-phase and Quadrature samples from Synchronization Signal Blocks and applies the Federated Averaging algorithm to a one-dimensional convolutional neural network. The approach reaches 97 percent accuracy and F1-score, matching or exceeding several centralized machine learning baselines. A sympathetic reader would care because it shows how 5G security can be added without creating a central repository of sensitive RF data that could itself become a target.

Core claim

The paper claims that a federated learning system using Federated Averaging to train a 1D convolutional neural network on over-the-air IQ samples extracted from Synchronization Signal Blocks achieves 97 percent accuracy and 97 percent F1-score for RF jamming detection. This performance surpasses centralized models including multilayer perceptron, 1D CNN, support vector machine, and logistic regression while ensuring that raw signal data never leaves the participating user equipment.

What carries the argument

Federated Averaging algorithm applied to a 1D convolutional neural network that learns from distributed IQ samples taken from Synchronization Signal Blocks, aggregating only model parameters instead of raw data.

If this is right

Jamming detection can be deployed across many user devices without creating a central database of raw RF signals.
The federated model matches or exceeds the detection rates of models trained on pooled centralized data.
The framework uses existing Synchronization Signal Blocks, so it requires no new signaling overhead for sample collection.
Privacy is maintained because only model updates are shared, reducing exposure of user equipment data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

In a real network the smaller size of model updates compared with raw IQ samples could reduce uplink traffic during training.
The same local-sample approach could be retrained on labeled data for other RF anomalies such as spoofing or interference without redesigning the collection pipeline.
Non-identical data distributions across user equipment might slow convergence and would need testing in heterogeneous deployments.

Load-bearing premise

The collected or simulated over-the-air IQ samples from Synchronization Signal Blocks are representative of real jamming attacks and that the federated averaging process adds acceptable communication and convergence overhead in an actual 5G network.

What would settle it

Evaluating the trained federated model on fresh over-the-air IQ samples captured from a live 5G base station during a controlled jamming attack and observing whether accuracy remains above 90 percent.

Figures

Figures reproduced from arXiv: 2605.01705 by Burak Kantarci, Mohammadreza Amini, Samhita Kuili.

**Figure 1.** Figure 1: System architecture III. SYSTEM MODEL We consider a federated learning (FL) framework deployed in a 5G and beyond wireless network consisting of a server S employed at gNB and a set of K UEs denoted as U = {UE1, UE2, . . . , UEK} shown in view at source ↗

**Figure 2.** Figure 2: Time-Frequency grid taken from the 5G operator in the absence of jamming signal. Detected SSBs Detected SSBs view at source ↗

**Figure 7.** Figure 7: Extracted SSB- jamming transmit gain is set to -50 dB. A. Federated learning Federated learning enables a single global model to train on data that remains on multiple separate UE in 5G network, often called clients. Each client is responsible to train on local dataset without sharing the raw data to the global model. FL instantiates an optimization problem that aims to minimize the global loss function ex… view at source ↗

**Figure 8.** Figure 8: Accuracy of FL with communication rounds view at source ↗

**Figure 9.** Figure 9: Loss of FL with communication rounds view at source ↗

read the original abstract

Jamming attacks are proliferating and pose a significant threat to the security of 5G and beyond networks. These attacks target 5G radio frequency (RF) domain and can disrupt the communication in wireless networks. While conventional machine learning and deep learning approaches demonstrate its potential for jamming detection, they typically require centralized data collection, compromising the privacy of user equipment (UEs). This work proposes a federated learning (FL)-based jamming detection framework that operates on over-the-air In-phase and Quadrature (IQ) samples extracted from Synchronization Signal Blocks (SSBs) in the RF domain. The framework enables collaborative model training across multiple UEs without sharing raw RF signal data. We adopt Federated Averaging (FedAvg) algorithm to train a 1D convolutional neural network (1DCNN) for effective detection of attacks. Numerical results demonstrate that the proposed FL framework achieves 97% accuracy and 97% F1-score, outperforming centralized baselines including MLP, 1DCNN, SVM, and logistic regression, while preserving the data privacy of all participating UEs

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The abstract reports strong FL performance for 5G jamming detection but lacks the details needed to trust the claim that it beats a properly trained centralized 1DCNN.

read the letter

The main takeaway here is that this work applies federated averaging to a 1D CNN for detecting RF jamming using IQ samples from 5G synchronization signal blocks, and it reports 97% accuracy while saying it beats centralized baselines. That combination of FL and this specific RF input for jamming seems new. The paper does a decent job laying out the privacy problem with centralized data collection for UE signals and shows how FL can address it without sharing raw data. Using over-the-air SSB samples is a practical choice for 5G networks. The soft spots are bigger. The abstract gives the 97% accuracy and F1 but supplies zero information on how much data was used, how the jamming was modeled or injected, what the train-test split looked like, or any hyperparameter settings. Without that, the numbers are hard to evaluate. The claim that the FL version outperforms the centralized 1DCNN is particularly shaky. Normally the centralized model with all the data should do at least as well, so either the centralized baseline was under-trained or there's something else going on with regularization or client data distribution. The stress-test note is right to flag this; the paper needs to show that the centralized run had equivalent total samples and training effort. No discussion of communication costs or convergence speed in a real network either, which matters for whether this is deployable. This paper is for researchers and engineers working on 5G security and applying distributed learning to wireless problems. A reader interested in privacy-preserving anomaly detection in RF could get some ideas from it, but only if the full paper has solid experiments. I think it deserves a serious referee to check the experimental details and see if the outperformance holds up. I'd recommend sending it to review rather than desk rejecting, with the expectation that the authors will need to add the missing setup information and perhaps rerun baselines fairly.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a federated learning (FL) framework for RF jamming detection in 5G networks. It extracts over-the-air In-phase/Quadrature (IQ) samples from Synchronization Signal Blocks (SSBs) and applies the Federated Averaging (FedAvg) algorithm to train a 1D convolutional neural network (1DCNN) collaboratively across user equipments (UEs) without sharing raw data. The central empirical claim is that the FL 1DCNN achieves 97% accuracy and 97% F1-score, outperforming centralized baselines (MLP, 1DCNN, SVM, logistic regression) while preserving UE data privacy.

Significance. If the performance comparison is shown to be robust, the work would provide concrete evidence that privacy-preserving FL can match or exceed centralized detection for RF-domain threats in 5G, a practically relevant result for secure wireless systems. The choice of SSB IQ samples as features is a strength, as it ties the method directly to standard 5G signaling.

major comments (2)

Numerical Results section: the headline claim that the proposed FL 1DCNN reaches 97% accuracy/F1 and outperforms the centralized 1DCNN baseline is load-bearing for the paper's contribution. The manuscript must explicitly state the total number of IQ samples used for the centralized 1DCNN, the number of epochs, optimizer settings, and learning-rate schedule, and confirm that these match the aggregate across all FL clients. If the centralized model was trained on a smaller subset or with mismatched hyperparameters, the reported outperformance is an artifact rather than evidence that FL is superior.
Experimental Setup section: no information is provided on dataset size, train-test split ratios, jamming signal models (e.g., power, bandwidth, duty cycle), or statistical significance testing (e.g., multiple random seeds, confidence intervals). Without these details the 97% figures cannot be reproduced or compared to prior work on RF jamming detection.

minor comments (2)

Abstract and Section 4: the phrase 'outperforming centralized baselines including MLP, 1DCNN, SVM, and logistic regression' should be accompanied by a table that reports all four metrics (accuracy, precision, recall, F1) for every method under identical conditions.
Figure captions and Section 3: clarify whether the 1DCNN architecture (number of layers, kernel sizes, pooling) is identical between the FL and centralized versions; any difference must be justified.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. The comments have prompted us to enhance the transparency of our experimental methodology and ensure the robustness of our comparative analysis. We address each major comment below and have made corresponding revisions to the paper.

read point-by-point responses

Referee: Numerical Results section: the headline claim that the proposed FL 1DCNN reaches 97% accuracy/F1 and outperforms the centralized 1DCNN baseline is load-bearing for the paper's contribution. The manuscript must explicitly state the total number of IQ samples used for the centralized 1DCNN, the number of epochs, optimizer settings, and learning-rate schedule, and confirm that these match the aggregate across all FL clients. If the centralized model was trained on a smaller subset or with mismatched hyperparameters, the reported outperformance is an artifact rather than evidence that FL is superior.

Authors: We fully agree that the comparison between FL and centralized learning must be conducted under identical conditions to be meaningful. Our centralized 1DCNN baseline was indeed trained on the complete aggregated dataset comprising all IQ samples from the participating UEs, using the same total sample count as the union of all FL clients' data. The training hyperparameters, including the number of epochs, optimizer, and learning rate schedule, were set identically to those used in the federated setting. In the revised version of the manuscript, we will add explicit statements and possibly a dedicated paragraph in the Numerical Results section to document these details and confirm the matching conditions. This will substantiate that the observed performance advantage is not due to any discrepancy in training setup. revision: yes
Referee: Experimental Setup section: no information is provided on dataset size, train-test split ratios, jamming signal models (e.g., power, bandwidth, duty cycle), or statistical significance testing (e.g., multiple random seeds, confidence intervals). Without these details the 97% figures cannot be reproduced or compared to prior work on RF jamming detection.

Authors: We acknowledge the need for more comprehensive information in the Experimental Setup section to facilitate reproducibility. In the revised manuscript, we will include the total dataset size, the specific train-test split ratios employed, detailed descriptions of the jamming signal models (such as their power levels, bandwidth, and duty cycles), and the procedures for statistical significance testing, including the use of multiple random seeds and the reporting of confidence intervals. These additions will allow readers to reproduce our 97% accuracy and F1-score results and compare them effectively with prior work. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance comparison is self-contained

full rationale

The paper presents an applied empirical study: it trains a 1DCNN via FedAvg on over-the-air IQ samples from SSBs, reports 97% accuracy/F1, and compares against centralized baselines (MLP, 1DCNN, SVM, logistic regression). No mathematical derivation, uniqueness theorem, or prediction is claimed that reduces by construction to fitted parameters, self-citations, or ansatz. The central claim is a direct experimental outcome on collected/simulated data; any concerns about training parity or data representativeness are validity issues, not circularity. The work contains no load-bearing self-referential steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work relies on standard assumptions of federated learning (IID or non-IID data across UEs, reliable uplink for model updates) and on the premise that SSB IQ samples contain sufficient discriminative features for jamming; no new entities or ad-hoc constants are introduced in the abstract.

axioms (1)

domain assumption Federated averaging converges to a useful global model when local datasets are drawn from similar distributions.
Implicit in the use of FedAvg for collaborative training across UEs.

pith-pipeline@v0.9.0 · 5506 in / 1225 out tokens · 36406 ms · 2026-05-09T17:00:36.056206+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 1 canonical work pages

[1]

Scalable cell-free massive mimo systems,

E. Bj ¨ornson and L. Sanguinetti, “Scalable cell-free massive mimo systems,”IEEE Transactions on Communications, vol. 68, no. 7, pp. 4247–4261, 2020

2020
[2]

A miniaturized microstrip antenna array at 5g millimeter-wave band,

X. Shen, Y . Liu, L. Zhao, G.-L. Huang, X. Shi, and Q. Huang, “A miniaturized microstrip antenna array at 5g millimeter-wave band,”IEEE Antennas and Wireless Propagation Letters, vol. 18, no. 8, pp. 1671– 1675, 2019

2019
[3]

Lte-advanced carrier aggregation for enhance- ment of bandwidth,

A. Goyal and K. Kumar, “Lte-advanced carrier aggregation for enhance- ment of bandwidth,” inAdvances in VLSI, Communication, and Signal Processing: Select Proceedings of VCAS 2018. Springer, 2019, pp. 341–351

2018
[4]

A compre- hensive survey on software-defined networking for smart communities,

R. Chaudhary, G. S. Aujla, N. Kumar, and P. K. Chouhan, “A compre- hensive survey on software-defined networking for smart communities,” Inl. J. of Communication Systems, vol. 38, no. 1, p. e5296, 2025

2025
[5]

A practical downlink noma scheme for wireless lans,

P. K. Sangdeh, H. Pirayesh, Q. Yan, K. Zeng, W. Lou, and H. Zeng, “A practical downlink noma scheme for wireless lans,”IEEE transactions on communications, vol. 68, no. 4, pp. 2236–2250, 2020

2020
[6]

A tutorial on beam management for 3gpp nr at mmwave frequencies,

M. Giordani, M. Polese, A. Roy, D. Castor, and M. Zorzi, “A tutorial on beam management for 3gpp nr at mmwave frequencies,”IEEE Communications Surveys & Tutorials, vol. 21, no. 1, pp. 173–196, 2018

2018
[7]

Jamming attacks and anti-jamming strategies in wireless networks: A comprehensive survey,

H. Pirayesh and H. Zeng, “Jamming attacks and anti-jamming strategies in wireless networks: A comprehensive survey,”IEEE communications surveys & tutorials, vol. 24, no. 2, pp. 767–809, 2022

2022
[8]

Federated learning- based jamming detection for distributed tactical wireless networks,

A. Meftah, G. Kaddoum, T. N. Do, and C. Talhi, “Federated learning- based jamming detection for distributed tactical wireless networks,” inMILCOM 2022-2022 IEEE Military Communications Conference (MILCOM). IEEE, 2022, pp. 629–634

2022
[9]

Real-world considerations for deep learning in wireless signal identification based on spectral correlation function,

K. Tekbıyık, ¨O. Akbunar, A. R. Ekti, A. G ¨orc ¸in, and G. K. Kurt, “Real-world considerations for deep learning in wireless signal identification based on spectral correlation function,”arXiv preprint arXiv:2003.08359, 2020

work page arXiv 2003
[10]

End-to-end learning from spectrum data: A deep learning approach for wireless signal identification in spectrum monitoring applications,

M. Kulin, T. Kazaz, I. Moerman, and E. De Poorter, “End-to-end learning from spectrum data: A deep learning approach for wireless signal identification in spectrum monitoring applications,”IEEE access, vol. 6, pp. 18 484–18 501, 2018

2018
[11]

Federated learning for 5g and beyond, a blessing and a curse- an experimental study on intrusion detection systems,

T. E. T. Djaidja, B. Brik, A. Boualouache, S. M. Senouci, and Y . Ghamri- Doudane, “Federated learning for 5g and beyond, a blessing and a curse- an experimental study on intrusion detection systems,”Computers & Security, vol. 139, p. 103707, 2024

2024
[12]

A privacy-preserving collaborative jamming attacks detection framework using federated learning,

Z. Abou El Houda, D. Naboulsi, and G. Kaddoum, “A privacy-preserving collaborative jamming attacks detection framework using federated learning,”IEEE Internet of Things Journal, vol. 11, no. 7, pp. 12 153– 12 164, 2023

2023
[13]

Federated learning-enabled jamming detection and waveform classification for distributed tactical wireless networks,

A. Meftah, T. N. Do, G. Kaddoum, C. Talhi, and S. Singh, “Federated learning-enabled jamming detection and waveform classification for distributed tactical wireless networks,”IEEE Transactions on Network and Service Management, vol. 20, no. 4, pp. 5053–5072, 2023

2023
[14]

Evaluation of na ¨ıve bayesian algorithms for cyber-attacks detection in wireless sensor networks,

S. Ismail and H. Reza, “Evaluation of na ¨ıve bayesian algorithms for cyber-attacks detection in wireless sensor networks,” in2022 IEEE world AI IoT congress (AIIoT). IEEE, 2022, pp. 283–289

2022
[15]

Multi-stage jamming attacks detection using deep learning combined with kernelized support vector machine in 5g cloud radio access networks,

M. Hachimi, G. Kaddoum, G. Gagnon, and P. Illy, “Multi-stage jamming attacks detection using deep learning combined with kernelized support vector machine in 5g cloud radio access networks,” inIntl. Symp. on networks, computers and communications. IEEE, 2020, pp. 1–5

2020
[16]

Federated-learning-based anomaly detection for iot security attacks,

V . Mothukuri, P. Khare, R. M. Parizi, S. Pouriyeh, A. Dehghantanha, and G. Srivastava, “Federated-learning-based anomaly detection for iot security attacks,”IEEE Internet of Things Journal, vol. 9, no. 4, pp. 2545–2554, 2021

2021
[17]

Intrusion detection for wireless edge networks based on federated learning,

Z. Chen, N. Lv, P. Liu, Y . Fang, K. Chen, and W. Pan, “Intrusion detection for wireless edge networks based on federated learning,”IEEE Access, vol. 8, pp. 217 463–217 472, 2020

2020
[18]

Federated learning for internet of things,

T. Zhang, C. He, T. Ma, L. Gao, M. Ma, and S. Avestimehr, “Federated learning for internet of things,” inProceedings of the 19th ACM conference on embedded networked sensor systems, 2021, pp. 413–419

2021
[19]

Decoding 5g security: toward a hybrid threat ontology,

R. A. Paskauskas, “Decoding 5g security: toward a hybrid threat ontology,”Open Research Europe, vol. 4, p. 34, 2025

2025
[20]

Survey on 5g physical layer security threats and countermeasures,

M. Harvanek, J. Bolcek, J. Kufa, L. Polak, M. Simka, and R. Marsalek, “Survey on 5g physical layer security threats and countermeasures,” Sensors, vol. 24, no. 17, p. 5523, 2024

2024
[21]

A formal analysis of 5g eap-tls protocol,

M. Shi, J. Chen, Z. Ma, K. He, M. Jia, and R. Du, “A formal analysis of 5g eap-tls protocol,”IEEE Transactions on Networking, 2025

2025
[22]

Toward an era of secure 5g convergence applications: Formal security verification of 3gpp akma with tls 1.3 psk option,

Y . Ko, I. W. A. J. Pawana, T. Won, P. V . Astillo, and I. You, “Toward an era of secure 5g convergence applications: Formal security verification of 3gpp akma with tls 1.3 psk option,”Applied Sciences, vol. 14, no. 23, p. 11152, 2024

2024
[23]

Synchronization procedure in 5G NR systems,

A. Omri, M. Shaqfeh, A. Ali, and H. Alnuweiri, “Synchronization procedure in 5G NR systems,”IEEE Access, vol. 7, pp. 41 286–41 295, 2019

2019
[24]

5G; NR; Physical channels and modulation,

3GPP, “5G; NR; Physical channels and modulation,” 3rd Generation Partnership Project (3GPP), Technical Specification (TS) 38.211, 10 2023, version 17.6.0. [Online]. Available: https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDet- ails.aspx?specificationId=3213

2023
[25]

Communication-efficient learning of deep networks from decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. Pmlr, 2017, pp. 1273– 1282

2017
[26]

Wsn-ds: a dataset for intrusion detection systems in wireless sensor networks,

I. Almomani, B. Al-Kasasbeh, and M. Al-Akhras, “Wsn-ds: a dataset for intrusion detection systems in wireless sensor networks,”Journal of Sensors, vol. 2016, no. 1, p. 4731953, 2016

2016
[27]

Deep learning and recognition of radar jamming based on cnn,

Q. Liu and W. Zhang, “Deep learning and recognition of radar jamming based on cnn,” in2019 12th international symposium on computational intelligence and design (ISCID), vol. 1. IEEE, 2019, pp. 208–212

2019
[28]

Iot-based smart farming architecture using federated learning: a nitrous oxide emission prediction use case,

P. Killeen, C. Lin, F. Li, I. Kiringa, and T. Yeap, “Iot-based smart farming architecture using federated learning: a nitrous oxide emission prediction use case,”ACM Journal on Computing and Sustainable Societies, vol. 3, no. 2, pp. 1–38, 2025

2025