pith. sign in

arxiv: 2604.11294 · v1 · submitted 2026-04-13 · 📡 eess.SP

Disaggregated multi-domain interference classification for O-RAN

Dieter Verbruggen , Hazem Sallouha , Sofie Pollin This is my paper

Pith reviewed 2026-05-10 15:32 UTC · model grok-4.3

classification 📡 eess.SP
keywords O-RANcross-technology interferencemulti-domain classificationdisaggregated architecturelow-latency inferencespectrum sharingsignal processing
0
0 comments X

The pith

A distributed multi-domain classifier for O-RAN interference runs in 400 microseconds on standard CPUs while cutting latency ninefold and computation elevenfold.

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

The paper establishes that O-RAN functional splits can be used to classify cross-technology interference by handling time, frequency, and channel state information representations locally in separate nodes before fusing the results. This distributed strategy avoids the high latency and compute demands of single monolithic models. A reader would care because dynamic spectrum sharing needs fast, efficient interference detection to keep quality of service without expensive hardware. The approach accepts a small accuracy cost in exchange for major gains in speed and efficiency that fit ordinary CPUs.

Core claim

The authors establish that a cost-aware multi-domain fusion architecture, where each domain is handled locally and then combined, delivers 400 μs inference latency on standard CPUs. This is nine times faster and eleven times cheaper computationally than a monolithic frequency-domain classifier, with only a 4 percent drop in accuracy and over 90 percent accuracy retained under high interference.

What carries the argument

The multi-domain distributed fusion architecture that processes raw time, frequency, and CSI signal representations locally within the O-RAN disaggregated nodes before combining results.

If this is right

  • Real-time control loops for interference mitigation become feasible on standard hardware in spectrum-sharing networks.
  • Computational load drops enough to support wider use of classification at edge nodes without specialized accelerators.
  • Classification accuracy stays above 90 percent even when interference levels rise.
  • The same local-domain strategy could be applied to other real-time O-RAN tasks such as channel estimation.

Where Pith is reading between the lines

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

  • Disaggregation of this kind could scale spectrum reuse in dense deployments by lowering the cost of continuous monitoring.
  • Similar splits might be tested for efficiency gains in related wireless tasks such as modulation recognition.
  • Adding domain-specific processing elements at each O-RAN node could push latency even lower in future hardware.

Load-bearing premise

The O-RAN functional split gives direct access to multi-domain raw signal data from the same stream without adding meaningful latency or losing key information.

What would settle it

An end-to-end test that measures the added latency when pulling time, frequency, and CSI data in parallel from an O-RAN split and checks whether accuracy drops below 85 percent in high-interference conditions.

Figures

Figures reproduced from arXiv: 2604.11294 by Dieter Verbruggen, Hazem Sallouha, Sofie Pollin.

Figure 1
Figure 1. Figure 1: System model of the proposed O-RAN-compliant uplink system, illustrating the extraction and processing of time, frequency, and CSI representations [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detailed view of the proposed OSIRIS architecture, including (a) the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Classification accuracy across varying SIR [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Classification accuracy versus SIR for (a) RADAR, (b) Wi-Fi, and (c) NR [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
read the original abstract

Spectrum sharing and dynamic spectrum reuse are becoming increasingly critical in modern wireless networks to address spectrum scarcity. However, these techniques inevitably increase Cross-Technology Interference (CTI). In this context, the Open Radio Access Network (O-RAN), as a modern and disaggregated network architecture, necessitates accurate, low-latency, and computationally efficient CTI classification and mitigation to support real-time control and maintain Quality of Service (QoS). Unfortunately, existing solutions predominantly rely on high-complexity, monolithic deep learning-based solutions that, while achieving high classification accuracy, incur significant latency and computational overhead This paper exploits the O-RAN functional split to leverage multi-domain raw signal representations (time, frequency, and Channel State Information (CSI)) directly from the same data stream. Each domain is processed locally, naturally interleaving CTI within the distributed, disaggregated O-RAN architecture. This distributed strategy enables a cost-aware, multi-domain fusion architecture that balances classification accuracy with computational overhead and latency. Our proposed multi-domain distributed architecture achieves a 400 $\mu s$ inference latency on standard CPUs. Compared to a state-of-the-art monolithic frequency-domain classifier, this represents an average 9x reduction in latency and an 11-fold decrease in computational cost, while sacrificing only 4% in classification performance and maintaining >90% accuracy in high-interference 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.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a disaggregated multi-domain architecture for cross-technology interference (CTI) classification in O-RAN. It exploits the O-RAN functional split to access and process time-domain, frequency-domain, and CSI representations locally from the same data stream, enabling a distributed, cost-aware fusion strategy. The central claim is that this yields 400 μs CPU inference latency, an average 9× latency reduction and 11× computational cost reduction versus a monolithic frequency-domain baseline, at the expense of only 4% classification accuracy while preserving >90% accuracy in high-interference regimes.

Significance. If the latency, cost, and accuracy claims are rigorously validated on compliant O-RAN stacks with realistic overheads, the work would offer a practically relevant advance for real-time interference management in disaggregated RANs. The alignment with O-RAN splits and emphasis on multi-domain local processing could inform efficient spectrum-sharing deployments; however, the significance is currently limited by the absence of detailed methodology, datasets, and overhead measurements needed to substantiate the headline numbers.

major comments (2)
  1. [Abstract] Abstract: The performance claims (400 μs latency, 9×/11× improvements, 4% accuracy sacrifice) are stated without any description of the experimental methodology, datasets, training procedures, hardware platform, or statistical tests. These details are required to assess whether the reported gains are reproducible or whether the comparison to the monolithic baseline is fair.
  2. [Abstract] Abstract and architecture description: The latency and cost reductions rest on the unquantified assumption that time-, frequency-, and CSI-domain representations are available locally from the identical data stream with negligible extra latency or information loss. O-RAN functional splits (e.g., 7.2x) impose fixed interfaces and data formats; any domain conversion, buffering, or inter-unit transfer would add overhead absent from a monolithic implementation and could materially reduce the claimed 9×/11× gains.
minor comments (2)
  1. [Abstract] The abstract refers to a 'cost-aware, multi-domain fusion architecture' without defining the fusion rule, the cost metric, or how the domains are weighted.
  2. [Abstract] No mention of the specific O-RAN split version or the exact interface points at which the multi-domain signals are extracted.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will revise the paper to improve clarity on experimental details and O-RAN assumptions while preserving the core contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The performance claims (400 μs latency, 9×/11× improvements, 4% accuracy sacrifice) are stated without any description of the experimental methodology, datasets, training procedures, hardware platform, or statistical tests. These details are required to assess whether the reported gains are reproducible or whether the comparison to the monolithic baseline is fair.

    Authors: We agree the abstract is concise and omits methodological context. The full manuscript details the evaluation in Section V, covering the CTI datasets (synthetic and over-the-air captures), training procedures (supervised learning with cross-entropy on a GPU cluster), hardware (standard Intel Xeon CPUs for inference), and statistical tests (multiple independent runs with 95% confidence intervals). The monolithic baseline uses identical data splits and metrics. We will revise the abstract to briefly reference the evaluation platform and hardware while directing readers to Section V for full reproducibility details. revision: yes

  2. Referee: [Abstract] Abstract and architecture description: The latency and cost reductions rest on the unquantified assumption that time-, frequency-, and CSI-domain representations are available locally from the identical data stream with negligible extra latency or information loss. O-RAN functional splits (e.g., 7.2x) impose fixed interfaces and data formats; any domain conversion, buffering, or inter-unit transfer would add overhead absent from a monolithic implementation and could materially reduce the claimed 9×/11× gains.

    Authors: We acknowledge this is a substantive point about practical deployment. Our design places domain-specific processing locally at the relevant O-RAN unit per the functional split (time-domain at RU, frequency-domain at DU), avoiding inter-unit transfers for the raw representations. However, explicit quantification of any buffering or format-conversion overheads is not provided in the current manuscript. We will expand Section III to discuss O-RAN split 7.2x interfaces, explain why overhead remains negligible under the standard, and add a qualitative analysis of potential impacts on the reported gains. revision: partial

standing simulated objections not resolved
  • Explicit numerical measurements of domain-conversion and inter-unit transfer overheads on a fully compliant O-RAN stack are absent from the manuscript and cannot be added without new experiments.

Circularity Check

0 steps flagged

Empirical performance claims rest on measured implementation, not self-referential derivation

full rationale

The paper reports measured results from a proposed multi-domain distributed CTI classifier implemented on standard CPUs, including 400 μs inference latency, 9x latency reduction, and 11x computational cost reduction versus an external state-of-the-art monolithic frequency-domain baseline, with only 4% accuracy sacrifice. These outcomes are presented as experimental outcomes of the disaggregated O-RAN architecture rather than outputs of any mathematical derivation, fitted parameter, or self-citation chain. No equations, uniqueness theorems, or ansatzes are invoked that reduce the claimed performance to the inputs by construction. The O-RAN split access is treated as an enabling architectural feature whose overhead is implicitly addressed by the reported measurements, not derived circularly from the model itself.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

With only the abstract available, no specific free parameters, axioms, or invented entities can be identified beyond standard domain assumptions in O-RAN and ML classification. The work relies on existing O-RAN architecture and multi-domain signal processing without introducing new entities or derivations.

axioms (1)
  • domain assumption O-RAN functional split enables local multi-domain processing from same data stream
    Invoked in the description of leveraging multi-domain representations directly from the same data stream.

pith-pipeline@v0.9.0 · 5541 in / 1307 out tokens · 34227 ms · 2026-05-10T15:32:46.417674+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

  1. [1]

    6g and beyond: The future of wireless communications systems,

    I. Akyildiz, A. Kak, and S. Nie, “6g and beyond: The future of wireless communications systems,”IEEE Access, vol. PP, pp. 1–1, 07 2020

    work page 2020

  2. [2]

    Coexistence analysis of wi-fi 6e and 5g nr-u in the 6 ghz band,

    N. Keshtiarast and M. Petrova, “Coexistence analysis of wi-fi 6e and 5g nr-u in the 6 ghz band,” inProceedings of the 2025 International Conference on Ns-3, ICNS3 ’25, (New York, NY , USA), p. 38–45, Association for Computing Machinery, 2025

    work page 2025

  3. [3]

    A machine learning-based approach for interference mitigation to enhance qos and qoe in 5g o-ran networks,

    D. Anand, M. A. Togou, and G.-M. Muntean, “A machine learning-based approach for interference mitigation to enhance qos and qoe in 5g o-ran networks,” in2024 IEEE 35th International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), pp. 1–6, 2024

    work page 2024

  4. [4]

    Un- derstanding o-ran: Architecture, interfaces, algorithms, security, and re- search challenges,

    M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Un- derstanding o-ran: Architecture, interfaces, algorithms, security, and re- search challenges,”IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376–1411, 2023

    work page 2023

  5. [5]

    O-RAN Fronthaul Working Group Control, User and Synchronization Plane Specification,

    O-RAN Alliance, “O-RAN Fronthaul Working Group Control, User and Synchronization Plane Specification,” technical specification, O-RAN Alliance, 2022. O-RAN.WG4.CUS.0-v08.00

    work page 2022

  6. [6]

    dapps: Enabling real-time ai-based open ran control,

    A. Lacava, L. Bonati, N. Mohamadi, R. Gangula, F. Kaltenberger, P. Johari, S. D’Oro, F. Cuomo, M. Polese, and T. Melodia, “dapps: Enabling real-time ai-based open ran control,”Computer Networks, vol. 269, p. 111342, 2025

    work page 2025

  7. [7]

    Dahlman, S

    E. Dahlman, S. Parkvall, and J. Skold,5G NR: The Next Generation Wireless Access Technology. Academic Press, 2020

    work page 2020

  8. [8]

    Interfo- ran: Real-time in-band cellular uplink interference detection with gpu- accelerated dapps,

    N. Neasamoni Santhi, D. Villa, M. Polese, and T. Melodia, “Interfo- ran: Real-time in-band cellular uplink interference detection with gpu- accelerated dapps,” inProceedings of the Twenty-Sixth International Symposium on Theory, Algorithmic F oundations, and Protocol Design for Mobile Networks and Mobile Computing, MobiHoc ’25, (New York, NY , USA), p. 71...

    work page 2025

  9. [9]

    Time–frequency fusion for enhancement of deep learning-based phys- ical layer identification,

    S. Zeng, Y . Chen, X. Li, J. Zhu, Y . Shen, and N. Shiratori, “Time–frequency fusion for enhancement of deep learning-based phys- ical layer identification,”Ad Hoc Networks, vol. 142, p. 103099, 2023

    work page 2023

  10. [10]

    Deep learning-based snr estimation,

    S. Zheng, S. Chen, T. Chen, Z. Yang, Z. Zhao, and X. Yang, “Deep learning-based snr estimation,”IEEE Open Journal of the Communica- tions Society, vol. 5, pp. 4778–4796, 2024

    work page 2024

  11. [11]

    Multi-domain networks for wireless interference recognition,

    P. Wang, Y . Cheng, B. Dong, Q. Peng, and S. Li, “Multi-domain networks for wireless interference recognition,”IEEE Transactions on V ehicular Technology, vol. 71, no. 6, pp. 6534–6547, 2022

    work page 2022

  12. [12]

    Ipfscnn: A time–frequency fusion cnn for wideband spectrum sensing,

    S.-Y . Kwon, D.-H. Park, and H.-N. Kim, “Ipfscnn: A time–frequency fusion cnn for wideband spectrum sensing,”Sensors, vol. 25, no. 23, 2025

    work page 2025

  13. [13]

    Multimodal machine learning: A survey and taxonomy,

    T. Baltru ˇsaitis, C. Ahuja, and L.-P. Morency, “Multimodal machine learning: A survey and taxonomy,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 41, no. 2, pp. 423–443, 2019

    work page 2019

  14. [14]

    Over the air deep learning based radio signal classification,

    T. O’Shea, T. Roy, and T. Clancy, “Over the air deep learning based radio signal classification,”IEEE Journal of Selected Topics in Signal Processing, vol. PP, 12 2017

    work page 2017