arxiv: 2604.15936 · v1 · submitted 2026-04-17 · 💻 cs.NI

Recognition: unknown

Federated Parameter-Efficient Adaptation for Interference Mitigation at the Wireless Edge

Evar Jones , Daniel J. Jakubisin , Sanmay Das

Authors on Pith no claims yet

Pith reviewed 2026-05-10 08:07 UTC · model grok-4.3

classification 💻 cs.NI

keywords federated learningparameter-efficient fine-tuningLoRAinterference mitigationtemporal convolutional networkswireless edge5G networksdistributed adaptation

0 comments

The pith

Placing low-rank adapters on frozen temporal CNNs and federating only those adapters enables interference mitigation across heterogeneous base stations with 20 times lower communication cost.

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

The paper establishes that low-rank adaptation can be inserted into the dilated layers of a temporal convolutional network used for suppressing co-channel interference. The backbone stays frozen to keep general signal extraction while the small adapters capture site-specific temporal patterns. These adapters, only 5.1 percent of the original parameters, are then averaged across nodes with FedAvg. A reader would care because full-model federated learning collapses under the non-IID interference seen in real deployments, yet this method still delivers roughly 12.6 percent average bit-error-rate reduction and helps nodes that have little local data. The result is a practical route to deploy learned interference mitigators without the communication scaling that dense networks cannot afford.

Core claim

By inserting low-rank adapters only into the dilated convolutional layers of the temporal CNN, the frozen backbone preserves shared signal-extraction features while the adapters learn local interference patterns; federating these adapters via FedAvg then yields 12.6 percent average BER improvement over the frozen model, nearly matching the 12.8 percent of purely local LoRA, outperforming local training on data-starved nodes, and avoiding the catastrophic failure of full-model FedAvg under heterogeneous conditions.

What carries the argument

Low-rank adapters placed on the dilated convolutional layers of a temporal CNN for interference suppression, aggregated across nodes by the standard FedAvg procedure.

If this is right

Communication volume per round falls by up to a factor of 20 relative to transmitting full model updates.
Nodes with few local samples receive performance gains from knowledge transferred through the federated adapters.
The method remains stable where full-model federated averaging produces sharp performance drops.
Local-only LoRA already improves BER by 12.8 percent on average; the federated version stays within 0.2 percent of that figure.

Where Pith is reading between the lines

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

The same placement of adapters on dilated layers could be tested on other temporal tasks such as channel prediction or beam tracking at the edge.
As network density grows, the 20-fold communication reduction would become increasingly decisive for feasibility.
If the backbone is pretrained on a wider range of interference types, the adapters might require even fewer local samples to reach target performance.

Load-bearing premise

The frozen backbone continues to extract useful general signal features even when interference statistics differ markedly from one base station to the next.

What would settle it

A controlled experiment in which base stations experience more extreme non-IID interference distributions than those simulated here, checking whether the federated adapters still deliver approximately 12 percent BER improvement on data-poor nodes.

Figures

Figures reproduced from arXiv: 2604.15936 by Daniel J. Jakubisin, Evar Jones, Sanmay Das.

**Figure 2.** Figure 2: Communication–performance tradeoff for federated methods. BER by 12.8% while training only 5.1% of the model parameters (14,400 vs. 281,954). Fed-LoRA performs comparably at 12.6%, confirming that federated aggregation of adapter parameters preserves most of the local adaptation quality. Both LoRA variants substantially outperform FiLM-based methods (L-FiLM: 6.3%, Fed-FiLM: 5.9%), which lack the capacity… view at source ↗

**Figure 3.** Figure 3: Per-interference-type BER on the global test set [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Per-node validation loss (MSE) across federated communication rounds. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

read the original abstract

Dense wireless deployments face co-channel interference from heterogeneous sources that vary across base stations (gNBs in 5G). While centralized DNN-based approaches to interference mitigation have shown strong performance, deploying and adapting these models across distributed gNBs via federated learning (FL) requires transmitting full model updates each round, resulting in a cost that scales poorly with network density. Parameter-efficient fine-tuning (PEFT) reduces this burden by training and communicating only a small fraction of parameters. While traditionally applied to large foundation models, we adapt Low-Rank Adaptation (LoRA) to temporal convolutional neural network architectures for interference suppression, placing low-rank adapters on the dilated convolutional layers. This placement enables LoRA to learn local interference-specific temporal patterns, while the frozen backbone retains the shared signal extraction capability. These lightweight adapters (5.1\% of backbone parameters) are federated via FedAvg, reducing per-round communication by up to 20$\times$ compared to federating full model updates. We evaluate various PEFT strategies across simulated distributed gNBs with non-IID interference environments. Results show that local LoRA achieves 12.8\% average BER improvement over the frozen backbone, while Fed-LoRA achieves comparable performance (12.6\%). Fed-LoRA outperforms local adaptation on data-starved nodes where federated knowledge transfer compensates for limited samples, all while avoiding the catastrophic degradation observed with full-model FedAvg under heterogeneous conditions.

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

This applies LoRA adapters on dilated layers of a TCN in a federated wireless setup and claims communication savings plus BER gains close to local adaptation, but the experiments lack the details needed to check the backbone assumption.

read the letter

The paper takes LoRA, places the adapters only on the dilated convolutional layers of a temporal conv net, and federates those adapters with FedAvg for interference mitigation across distributed gNBs. That specific placement and the resulting 20x communication cut while matching local LoRA performance (12.6% vs 12.8% BER improvement) is the concrete new piece. It also shows the federated version helping data-starved nodes without the collapse that full-model FedAvg shows under non-IID interference. Those are practical points for anyone already working on edge adaptation in dense wireless networks. The 5.1% parameter fraction is a clear win on the comms side and the abstract states the numbers directly. The central assumption is that the frozen backbone keeps extracting shared signal features even when interference patterns differ across nodes. The stress-test note is right to flag this: the abstract gives no backbone-only BER per node, no pre-training distribution details, and no ablation that isolates whether the adapters are really carrying the federated knowledge or whether the backbone is already degrading. Without those, the claim that federated adapter transfer is what prevents collapse rests on unverified ground. Simulation setup, node count, exact non-IID construction, error bars, and significance tests are also missing from the provided text, so the quantitative claims cannot be checked. This is a straightforward empirical extension of known PEFT and FL techniques rather than a new method. It is aimed at researchers in wireless ML and 5G/6G interference management who already follow federated adaptation work. The idea is clear enough and the communication angle is useful, so it deserves a serious referee to verify the experimental setup and test the backbone assumption with proper controls. I would not cite it yet but would want to see the full results before deciding.

Referee Report

2 major / 1 minor

Summary. The paper proposes adapting Low-Rank Adaptation (LoRA) to temporal convolutional networks for federated interference mitigation at distributed 5G gNBs. By placing low-rank adapters only on dilated convolutional layers (5.1% of backbone parameters), the approach freezes the backbone to retain shared signal extraction while federating adapters via FedAvg. This yields up to 20× communication reduction per round. Empirical results claim local LoRA achieves 12.8% average BER improvement over the frozen backbone, Fed-LoRA matches it at 12.6%, outperforms local adaptation on data-starved nodes, and avoids the degradation of full-model FedAvg under non-IID heterogeneous interference.

Significance. If the results hold with proper verification, the work demonstrates a practical application of parameter-efficient fine-tuning to smaller domain-specific models in wireless networks, enabling scalable federated adaptation without full-model transmission costs. The avoidance of catastrophic degradation under heterogeneity and benefits for data-starved nodes highlight potential for real-world edge deployment in dense 5G scenarios. It extends PEFT techniques beyond large foundation models to temporal CNNs for interference suppression.

major comments (2)

Abstract: The reported 12.8% (local LoRA) and 12.6% (Fed-LoRA) average BER improvements over the frozen backbone are presented without any details on simulation setup (number of gNBs, exact non-IID interference distributions, samples per node, error bars, or statistical significance tests). These omissions make the central performance claims unverifiable and prevent assessment of whether the gains are robust or attributable to the proposed method.
Abstract: The key assumption that the frozen backbone retains shared signal extraction capability across heterogeneous non-IID interference environments at distributed gNBs is stated as enabling the LoRA placement and Fed-LoRA success, but no supporting evidence is provided (e.g., per-gNB backbone-only BER results, ablation isolating backbone performance, or pre-training distribution details). This assumption is load-bearing for attributing comparable performance and avoidance of FedAvg collapse to federated adapter transfer.

minor comments (1)

Abstract: The evaluation of 'various PEFT strategies' is mentioned but not enumerated or compared in detail beyond LoRA and full FedAvg; a table or section listing all tested methods and their communication/performance trade-offs would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major point below with clarifications from the full paper and propose targeted revisions to enhance verifiability and strengthen the presentation of our assumptions.

read point-by-point responses

Referee: Abstract: The reported 12.8% (local LoRA) and 12.6% (Fed-LoRA) average BER improvements over the frozen backbone are presented without any details on simulation setup (number of gNBs, exact non-IID interference distributions, samples per node, error bars, or statistical significance tests). These omissions make the central performance claims unverifiable and prevent assessment of whether the gains are robust or attributable to the proposed method.

Authors: We agree that the abstract, by design, is a concise summary and omits granular experimental parameters. The full manuscript provides these details in Section 4 (Experimental Setup) and Section 5 (Results), including the distributed gNB configuration, modeling of non-IID interference, per-node sample counts, error bars from repeated trials, and statistical significance testing. To improve immediate verifiability, we will revise the abstract to include a brief parenthetical summary of the setup and analysis approach without exceeding typical length constraints. revision: partial
Referee: Abstract: The key assumption that the frozen backbone retains shared signal extraction capability across heterogeneous non-IID interference environments at distributed gNBs is stated as enabling the LoRA placement and Fed-LoRA success, but no supporting evidence is provided (e.g., per-gNB backbone-only BER results, ablation isolating backbone performance, or pre-training distribution details). This assumption is load-bearing for attributing comparable performance and avoidance of FedAvg collapse to federated adapter transfer.

Authors: The manuscript presents indirect support for this assumption through the main results: local LoRA and Fed-LoRA achieve comparable BER gains while full-model FedAvg exhibits catastrophic degradation under the same non-IID conditions, consistent with the backbone preserving transferable signal features. However, we acknowledge that explicit per-gNB backbone-only evaluations and pre-training details would make the claim more robust. We will add a dedicated ablation subsection with these analyses in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical performance evaluation

full rationale

The paper reports simulation results comparing local LoRA, Fed-LoRA, and full-model FedAvg on BER for interference mitigation using TCN backbones. No equations, derivations, or first-principles predictions are present. Claims rest on observed average improvements (12.8% local, 12.6% federated) and qualitative statements about adapter placement, none of which reduce to fitted inputs or self-citations by construction. The frozen-backbone assumption is an empirical premise tested via the reported experiments rather than a definitional or fitted tautology.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Paper is purely empirical and relies on standard assumptions from federated learning and neural network training literature; no free parameters, axioms, or invented entities are introduced beyond those implicit in LoRA and FedAvg.

pith-pipeline@v0.9.0 · 5559 in / 1148 out tokens · 25635 ms · 2026-05-10T08:07:06.053202+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

33 extracted references · 22 canonical work pages · 1 internal anchor

[1]

Bouziane Brik, Hatim Chergui, Lanfranco Zanzi, Francesco Devoti, Adlen Ksentini, Muhammad Shuaib Siddiqui, Xavier Costa-Pérez, and Christos Verikoukis. 2025. Explainable AI in 6G O-RAN: A Tutorial and Survey on Architecture, Use Cases, Challenges, and Future Research. IEEE Communications Surveys & Tutorials27, 5 (2025), 2826–. https: //doi.org/10.1109/COM...

work page doi:10.1109/comst.2024.3510543 2025
[2]

Chuntao Ding, Xu Cao, Jianhang Xie, Linlin Fan, Shangguang Wang, and Zhichao Lu. 2024. LoRA-C: Parameter-Efficient Fine-Tuning of Robust CNN for IoT Devices.arXiv preprint arXiv:2410.16954(2024)

work page arXiv 2024
[3]

Elbir and Sinem Coleri

Ahmet M. Elbir and Sinem Coleri. 2021. Federated Learning for Chan- nel Estimation in Conventional and RIS-Assisted Massive MIMO.IEEE Transactions on Wireless Communications20, 10 (2021), 6518–6533. https://doi.org/10.1109/TWC.2021.3081836

work page doi:10.1109/twc.2021.3081836 2021
[4]

Zhidong Gao, Zhenxiao Zhang, Yuanxiong Guo, and Yanmin Gong
[5]

InIEEE INFOCOM 2025 – IEEE 10 Conference on Computer Communications

Federated Adaptive Fine-Tuning of Large Language Models with Heterogeneous Quantization and LoRA. InIEEE INFOCOM 2025 – IEEE 10 Conference on Computer Communications. IEEE

2025
[6]

Zhiheng Guo, Zhaoyang Liu, Zihan Cen, Chenyuan Feng, Xinghua Sun, Xiang Chen, Tony Q. S. Quek, and Xijun Wang. 2025. CoCo-Fed: A Unified Framework for Memory- and Communication-Efficient Fed- erated Learning at the Wireless Edge.arXiv preprint arXiv:2601.00549 (2025)

work page arXiv 2025
[7]

Zhen Han, Ning Ding, Yuxian Wu, Xiang Zhao, Kai Lv, Yixuan Gu, Zhiyuan Liu, and Maosong Sun. 2024. Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey.Comput. Surveys(2024). https://doi.org/10.1145/3638552

work page doi:10.1145/3638552 2024
[8]

Lukas Henneke. 2024. Improving Data-Driven RF Signal Separation with SOI-Matched Autoencoders. InProceedings of the IEEE Interna- tional Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 45–46. https://doi.org/10.1109/ICASSPW62465.2024. 10626245

work page doi:10.1109/icasspw62465.2024 2024
[9]

Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Syl- vain Gelly. 2019. Parameter-Efficient Transfer Learning for NLP. In Proceedings of the 36th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 97). 2790–2799

2019
[10]

Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations (ICLR)

2022
[11]

Michael Buehrer

Hiten Prakash Kothari and R. Michael Buehrer. 2025. Interference Mitigation using U-Net Autoencoder based system.arXiv preprint arXiv:2512.13844(2025)

work page arXiv 2025
[12]

Alejandro Lancho, Amir Weiss, Gary C. F. Lee, Tejas Jayashankar, Binoy G. Kurien, Yury Polyanskiy, and Gregory W. Wornell. 2025. RF Challenge: The Data-Driven Radio Frequency Signal Separation Challenge.IEEE Open Journal of the Communications Society6 (2025), 4083–4100. https://doi.org/10.1109/OJCOMS.2025.3556319

work page doi:10.1109/ojcoms.2025.3556319 2025
[13]

H Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. InProceedings of the 20th In- ternational Conference on Artificial Intelligence and Statistics (AISTATS), Vol. 54. 1–10

2017
[14]

Mostafa Naseri, Jaron Fontaine, Ingrid Moerman, Eli De Poorter, and Adnan Shahid. 2024. A U-Net Architecture for Time-Frequency In- terference Signal Separation of RF Waveforms. InProceedings of the IEEE International Conference on Acoustics, Speech, and Signal Pro- cessing Workshops (ICASSPW). IEEE, 91–92. https://doi.org/10.1109/ ICASSPW62465.2024.10626228

work page arXiv 2024
[15]

Vincent Poor, and Adnan Shahid

Mostafa Naseri, Eli De Poorter, Ingrid Moerman, H. Vincent Poor, and Adnan Shahid. 2025. High-Throughput Adaptive Co-Channel Interference Cancellation for Edge Devices Using Depthwise Separa- ble Convolutions, Quantization, and Pruning.IEEE Open Journal of the Communications Society6 (2025), 656–. https://doi.org/10.1109/ OJCOMS.2024.3523797

work page arXiv 2025
[16]

Anselme Ndikumana, Kim Khoa Nguyen, and Mohamed Cheriet. 2023. Federated Learning Assisted Deep Q-Learning for Joint Task Offload- ing and Fronthaul Segment Routing in Open RAN.IEEE Transac- tions on Network and Service Management20, 3 (2023), 3261–3273. https://doi.org/10.1109/TNSM.2023.3245544

work page doi:10.1109/tnsm.2023.3245544 2023
[17]

Shah, Daniel J

Taiwo Oyedare, Vijay K. Shah, Daniel J. Jakubisin, and Jeffrey H. Reed
[18]

https://doi.org/10.1109/ACCESS.2022.3185124

Interference Suppression Using Deep Learning: Current Ap- proaches and Open Challenges.IEEE Access10 (2022), 66238–66266. https://doi.org/10.1109/ACCESS.2022.3185124

work page doi:10.1109/access.2022.3185124 2022
[19]

Marina, and Thanos Triantafyllou

Qingrui Pan, Mahesh K. Marina, and Thanos Triantafyllou. 2025. On NextG Open RAN as a Sensing Infrastructure. InProceedings of the 26th International Workshop on Mobile Computing Systems and Applications (HotMobile ’25). ACM, CA, La Quinta, USA. https://doi.org/10.1145/ 3708468.3711894

work page arXiv 2025
[20]

Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. 2018. FiLM: Visual Reasoning with a General Condi- tioning Layer. InProceedings of the Thirty-Second AAAI Conference on Artificial Intelligence. 3942–3951

2018
[21]

Ratun Rahman and Dinh C. Nguyen. 2024. Improved Modulation Recognition Using Personalized Federated Learning.IEEE Transactions on Vehicular Technology73, 12 (2024), 19937–19952. https://doi.org/10. 1109/TVT.2024.3447569

work page arXiv 2024
[22]

Dario Rethage, Jordi Pons, and Xavier Serra. 2018. A WaveNet for Speech Denoising.arXiv preprint arXiv:1706.07162(2018)

work page arXiv 2018
[23]

Yasintha Rumesh, Dinaj Attanayaka, Pawani Porambage, Jarno Pinola, Joshua Groen, and Kaushik Chowdhury. 2024. Federated Learning for Anomaly Detection in Open RAN: Security Architecture Within a Digital Twin. InProceedings of the European Conference on Networks and Communications and 6G Summit (EuCNC/6G Summit). 1–6. https: //doi.org/10.1109/EuCNC/6GSummi...

work page doi:10.1109/eucnc/6gsummit60053.2024.10597083 2024
[24]

Aliaksandra Shysheya, John Bronskill, Massimiliano Patacchiola, Se- bastian Nowozin, and Richard E. Turner. 2023. FiT: Parameter-Efficient Few-Shot Transfer Learning for Personalized and Federated Image Classification. InProceedings of the International Conference on Learn- ing Representations

2023
[25]

Amardip Kumar Singh and Kim Khoa Nguyen. 2024. Communication Efficient Compressed and Accelerated Federated Learning in Open RAN Intelligent Controllers.IEEE/ACM Transactions on Networking 32, 4 (2024), 3361–3376. https://doi.org/10.1109/TNET.2024.3384839

work page doi:10.1109/tnet.2024.3384839 2024
[26]

Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. 2024. Improving LoRA in Privacy-preserving Federated Learning. InProc. ICLR

2024
[27]

Yu Tian, Ahmed Alhammadi, Abdullah Quran, and Abubakar Sani Ali. 2024. A Novel Approach to WaveNet Architecture for RF Sig- nal Separation with Learnable Dilation and Data Augmentation. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 79–80. https: //doi.org/10.1109/ICASSPW62465.202...

work page doi:10.1109/icasspw62465.2024.10627586 2024
[28]

Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. 2016. WaveNet: A Generative Model for Raw Audio.arXiv preprint arXiv:1609.03499(2016)

work page internal anchor Pith review arXiv 2016
[29]

Shuo Wang, Tianxin Wang, and Xudong Wang. 2025. FedPDA: Col- laborative Learning for Reducing Online-Adaptation Frequency of Neural Receivers. InProceedings of the IEEE International Conference on Computer Communications (INFOCOM). 1–10. https://doi.org/10. 1109/INFOCOM55648.2025.11044747

work page arXiv 2025
[30]

Han Zhang, Hao Zhou, and Melike Erol-Kantarci. 2022. Federated Deep Reinforcement Learning for Resource Allocation in O-RAN Slicing. In2022 IEEE Global Communications Conference (GLOBECOM). IEEE, 958–963. https://doi.org/10.1109/GLOBECOM48099.2022.10001658

work page doi:10.1109/globecom48099.2022.10001658 2022
[31]

Zhongyuan Zhao, Chenyuan Feng, Wei Hong, Jiamo Jiang, Chao Jia, Tony Q. S. Quek, and Mugen Peng. 2022. Federated Learning With Non-IID Data in Wireless Networks.IEEE Transactions on Wireless Communications21, 3 (2022), 1927–1940. https://doi.org/10.1109/TWC. 2021.3108197

work page doi:10.1109/twc 2022
[32]

Zihan Zhong, Zhiqiang Tang, Tong He, Haoyang Fang, and Chun Yuan
[33]

Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model. InProc. ICLR. 11