Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Radio-FM: A Foundation Model for Radio Signal Representation Learning and Its Applications

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Radio-FM, a family of Transformer foundation models pretrained on raw I/Q radio signals with dual-channel attention, claims state-of-the-art transfer on 13 of 15 benchmarks across modulation, radar, emitter, wireless-technology, and…

desk verdict New architecture, but the headline transfer claims do not survive contact with the benchmark setup: 8 of 15 downstream datasets are also pretraining corpora. read the letter →

arxiv 2608.05793 v1 pith:P4IA3FKL submitted 2026-08-06 eess.SP

Jinchao Zhou , Wupeng Xie , Zhuangzhi Chen , Yao Lu , Qi Xuan , Yun Lin , Guan Gui This is my paper · ORCID
classification eess.SP
keywords radiofoundationmodelself-supervisedlearningI/Qsignalrepresentationmaskedautoencoderautomaticmodulationrecognitionspecificemitteridentificationfew-shottransfer
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 is trying to establish that one self-supervised backbone can serve as a general-purpose representation learner for radio-frequency signals in their native I/Q form, replacing task-specific models for modulation, radar, emitter, wireless-technology, and interference recognition. It argues that the key is to treat the in-phase and quadrature channels as independent streams processed by intra-channel temporal attention, with a lightweight inter-channel attention to model their mutual dependence, and to pretrain with channel-independent masked reconstruction on a heterogeneous corpus. On 15 downstream benchmarks, it reports state-of-the-art accuracy on 13, and in 10-, 50-, and 100-shot fine-tuning it reports large margins over existing radio foundation models on data-hungry tasks like radar-waveform and emitter identification. The broader point is that raw I/Q data, despite its heterogeneity, can support a single scalable pretrained backbone with emergent structure in unsupervised features.

What carries the argument

The load-bearing object is a dual-level Transformer encoder over tokenized I/Q waveforms. Each channel is tokenized by a 1D convolution with patch size 8, and each encoder layer runs intra-channel multi-head self-attention with rotary position embeddings across the temporal dimension, then pools each channel's tokens globally and runs inter-channel self-attention over the two channels to recalibrate every time step with the other channel's information. Pretraining uses channel-independent random masking at 60%, where I and Q masks are sampled separately, with a masked-reconstruction mean-squared-error objective, and masking preserves original patch indices so positional embeddings reflect the unmasked grid. Training on heterogeneous multi-source corpora is made scalable by a token-budgeted dynamic batching mechanism that sets the batch size from a per-device token budget rather than a fixed instance count.

What would settle it

Train a randomly initialized Radio-FM of the same architecture on each downstream task under identical fine-tuning settings; on any benchmark where from-scratch accuracy matches the pretrained-then-fine-tuned accuracy, the reported benefit of pretraining is not the source of the gain for that task.

Watch

Extended reading notes

Core claim

The central claim is that Radio-FM learns transferable radio-signal representations by processing the in-phase (I) and quadrature (Q) channels separately for temporal modeling, then combining them through a lightweight attention mechanism, and by pretraining with a masked-autoencoding objective where masks are sampled independently for the two channels. The paper reports that this design outperforms the previous radio foundation models it compares against, SpectrumFM and EMind, on 13 of 15 full fine-tuning benchmarks, and shows superior few-shot transferability: the XLarge variant is best on 9 of 15 datasets at 10 shots, and it leads by over 20 percentage points on DeepRadar2022. The claim also includes a scaling law in which larger variants improve data-scarce and high-complexity tasks while full-data fine-tuning saturates near the Base scale, and ablations attribute the gains specifically to channel-independent masking, inter-channel interaction, and layer scaling.

Load-bearing premise

The central results are transfer results, so they stand or fall on the assumption that the downstream evaluation sets were not part of the pretraining corpus; the paper lists the same dataset names in both its pretraining and downstream tables and never states a split that guarantees disjointness.

Editorial extensions

If this is right

  • A single Radio-FM backbone can replace separate supervised models for automatic modulation recognition, radar waveform classification, specific emitter identification, wireless technology recognition, and wireless interference identification; full fine-tuning improves on or matches the two baselines on 13 of 15 benchmarks.
  • In label-scarce regimes, the pretrained representations are the main driver of accuracy: the XLarge variant wins 9 of 15 datasets at 10-shot, and its margin over SpectrumFM on radar waveform recognition exceeds 20 percentage points.
  • Scaling from Tiny (2.38M parameters) to XLarge (114.64M parameters) produces consistent accuracy growth in few-shot and from-scratch settings, while full fine-tuning saturates near Base scale, indicating that larger models matter most when data is scarce or classes are fine-grained.
  • Channel-independent masking is the largest single design choice: removing it costs about 2 percentage points in accuracy, showing that forcing each channel to be reconstructed from the other teaches phase-aware representations that shared masking does not.

Reading between the lines

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

  • An implication the paper leaves implicit is that the same dual-channel, channel-independent masking recipe could transfer to other physical-layer two-stream inputs, such as polarimetric radar returns or multi-antenna I/Q pairs, where within-stream temporal structure and cross-stream correlation both carry information.
  • Because the few-shot protocol samples K instances per SNR level on the same benchmark datasets used for full evaluation, a stronger test of the transfer claim would hold out entire SNR bands or receiver conditions during fine-tuning; the paper does not report this, and it would directly probe the stated motivation of robustness under channel variation.
  • A practical extension would be releasing the pretrained encoders along with the standardization preprocessing, so downstream users could measure whether the reported margins survive on their own receiver hardware, which the paper does not evaluate.
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.

Referee Report

4 major / 5 minor

Summary. The paper introduces Radio-FM, a family of Transformer-based foundation models for raw I/Q radio signals. The method combines dual-channel patch embedding with intra-channel temporal self-attention, inter-channel attention, channel-independent masked reconstruction, and token-budgeted dynamic batching for heterogeneous multi-dataset pretraining. The authors pretrain five model variants (Tiny to XLarge) on 15 datasets and evaluate them on 15 downstream benchmarks covering modulation recognition, radar waveform classification, emitter identification, wireless technology recognition, and interference identification. The central claim is that Radio-FM achieves state-of-the-art performance on 13 of 15 benchmarks and exhibits superior few-shot transferability compared with SpectrumFM and EMind.

Significance. If the empirical claims were sound, Radio-FM would be a useful contribution to radio-signal foundation models: the dual-channel I/Q architecture, channel-independent masking, and dynamic batching recipe are sensible and the scaling study across five model sizes adds useful evidence. The paper is also commendable for standardizing a heterogeneous corpus and for including ablations on mask ratio, patch size, and architectural components. However, the headline claims of transfer and state-of-the-art performance are not supported by the experiments as presented: most downstream benchmarks overlap with the pretraining corpus, and the main results tables contain contradictions. The significance of the contribution therefore remains conditional on a clean, disjoint evaluation and corrected reporting.

major comments (4)
  1. [§IV.A.1, Tables I–II, Eq. (1)] The central transfer claim is undermined by benchmark leakage between pretraining and downstream evaluation. Table I lists RML2018.01a, Sig2019-12, HKDD AMC12, HKDD AMC36, HisarMod2019.1, DeepRadar2022, SubGHz, and TechRec as pretraining datasets, and Table II lists the same datasets as downstream benchmarks; WiSig-ManyTX is a split of pretraining's WiSig-Full. Section IV.A.1 states only that official train/test splits are used, and it never states that the pretraining corpus excluded these datasets or their train portions. Under Eq. (1), the encoder is optimized on unlabeled data from all 15 corpora, so fine-tuning and few-shot results on these eight benchmarks reflect within-corpus adaptation rather than transfer to new environments. This pattern is corroborated by Table IV: Radio-FM's full-fine-tuning misses (RML2016.10a, RML2022) are on datasets not in the pretraining corpus, while nearly every in-corpus benchmark is won. The paper must either pretrain on a corpus disjoint from all downstream benchmarks or clearly re-frame the in-corpus results as within-corpus adaptation and restrict the transfer claim to the non-overlapping datasets.
  2. [Tables V and VI] Table VI contradicts Table V in multiple load-bearing places. For example, HKDD AMC36 Base is reported as 56.27 in Table V (100-shot) but 51.16 in Table VI (Few-shot FT 100-shot); HisarMod2019.1 Large is 58.13 in Table V but 80.69 in Table VI. In addition, Table VI reports many 'From Scratch' values that exactly equal their 'Few-shot FT (100-shot)' counterparts (e.g., HKDD AMC36 Base 51.16/51.16, HisarMod2019.1 Large 80.69/80.69, RadChar-Base Base 89.55/89.55, DeepRadar2022 Base 76.07/76.07, ADSB-100 Base 57.30/57.30, SubGHz Base 83.78/83.78). Identical numbers across different training paradigms are not plausible and indicate a data-reporting error. Because these tables are the primary evidence for the few-shot transferability and scaling claims, the experimental record must be corrected and re-verified before the headline claims can be assessed.
  3. [§IV.B and Table IV] The claim of 'state-of-the-art performance on 13 of 15 datasets' is not supported by the baseline selection. The comparison is limited to SpectrumFM and EMind, which are two foundation models; no task-specific state-of-the-art methods for AMR, RWC, SEI, WTR, or WII are included. Many of these benchmarks have published specialized baselines with strong results, so beating two generalist backbones does not establish state of the art. The authors should either compare against representative task-specific methods or weaken the claim to 'improvement over the tested foundation models.'
  4. [§IV.A.3 and §III.D.1] There is an internal inconsistency in the masking ratio specification. Section III.D.1 states that a high masking ratio (e.g., 75%) is applied, while Section IV.A.3 reports that channel-independent masking is applied with a 60% ratio, and Fig. 9b identifies 60% as optimal. This is not fatal by itself, but the pretraining setup should state the exact ratio used for the main results.
minor comments (5)
  1. [Abstract, Introduction, §IV.B] The number of claimed state-of-the-art results is inconsistent: the Abstract and Section IV.B say 13 of 15, while the Introduction lists 'state-of-the-art performance on 12 tasks.' Please harmonize these numbers.
  2. [Fig. 1 caption] The caption 'AI-generated image using Nano Banana' is inappropriate for a scientific paper; a schematic diagram or a properly credited real system figure would be preferable.
  3. [§IV.C.1] There is a typo in 'prerained feature projections'; it should be 'pretrained.'
  4. [Table V] The 'Best Count All Tasks' row is difficult to read and appears to contain formatting errors (e.g., '0 19', '210'). Please format it as a clean table of counts per column.
  5. [Table III and Fig. 9] FLOPs are reported at a fixed sequence length of 1024, but many datasets have different lengths (e.g., 128, 256, 2048, 4096). Please clarify how the FLOPs values in Fig. 9 and Table III translate to the actual pretraining and downstream inputs.

Circularity Check

1 steps flagged · score 8.0 of 10

Downstream evaluation is circular: Table II benchmarks overlap Table I pretraining corpora, so fine-tuning and few-shot 'transfer' results are within-corpus adaptation.

  1. fitted input called prediction [Section IV.A.1 (Datasets), Tables I and II, Eq. (1)]
    "The pretraining phase utilizes 15 diverse datasets spanning modulation, radar, and communication signals (Fig. 4 and Table I). For downstream tasks, we evaluate on 15 datasets covering Automatic Modulation Recognition (AMR), Radio Waveform Classification (RWC), Specific Emitter Identification (SEI), and Wireless Technology Recognition (WTR), as detailed in Table II. Regarding data splitting, we prioritize the official training/test sets provided by the original benchmarks to ensure fair comparison."

    Table I lists RML2018.01a, Sig2019-12, HKDD AMC12/36, HisarMod2019.1, DeepRadar2022, WiSig-Full, SubGHz, and TechRec as pretraining corpora, and Table II lists the same dataset names as downstream benchmarks (WiSig-ManyTX comes from the same WiSig collection [25]). Eq. (1) optimizes the encoder on unlabeled tokens from these corpora before any downstream head is trained, and no exclusion or disjoint split is stated. Full fine-tuning and few-shot evaluation on these benchmarks therefore measure within-corpus adaptation to data whose distribution (and possibly test files) the encoder has already seen, not transfer to new environments.

full rationale

The core circularity is in the evaluation design, not in the architecture or pretraining recipe. Section IV.A.1 states that pretraining uses the 15 corpora of Table I and downstream evaluation uses Table II, with official splits, and it never states that pretraining excluded the downstream datasets or their test portions. Eight exact dataset names appear in both tables, and WiSig-ManyTX is drawn from the same WiSig collection as pretraining's WiSig-Full. Since Eq. (1) fits the encoder by masked reconstruction on unlabeled tokens from these corpora, the reported fine-tuning and few-shot gains on those benchmarks are within-corpus results rather than independent transfer. The remaining seven datasets (RML2016.10a/b, RML2022, RadChar-Base, ADSB-100, EM-Infer-Comm) are disjoint and provide some genuine evidence, and the dual-channel architecture and dynamic batching contributions are not circular. Self-citations in the related-work section are not load-bearing. I also flag, without classifying it as circularity, an internal reliability problem: Table VI's Few-shot FT (100-shot) values contradict Table V (e.g., HKDD AMC36 Base 51.16 vs 56.27; HisarMod2019.1 Large 80.69 vs 58.13) and several equal the From Scratch values, so the few-shot claims need verification independent of the overlap issue. Overall, the headline empirical result is largely forced by the overlapping pretraining and evaluation corpora, giving a circularity score of 8.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The model introduces no new physical entities. The free parameters are pretraining hyperparameters tuned by ablation. The key unstated premise is the disjointness of pretraining and downstream splits, which is violated by the listed tables.

free parameters (3)
  • mask ratio = 60%
    Chosen as optimal in the Fig. 9b ablation and fixed across all pretraining; the choice directly affects reconstruction difficulty and downstream transfer.
  • patch size = 8
    Chosen as optimal in the Fig. 9c ablation and fixed at P=8; this controls temporal resolution and model FLOPs.
  • token budget per device Cmax
    Introduced in Eq. 11 for dynamic batching, but no value or sensitivity analysis is reported; it governs batch composition under length heterogeneity.
assumptions (5)
  • standard math Transformer attention, RoPE, LayerScale, and DropPath behave as specified in the cited literature.
    Used without proof throughout Section III.C.
  • domain assumption Raw I/Q signals are best represented by processing I and Q channels independently with cross-channel attention.
    This is the core architectural premise, stated in Section III.C.
  • domain assumption Masked reconstruction pretraining on unlabeled I/Q transfers to downstream classification tasks.
    Adopted from MAE-style methods in Section III.A.
  • domain assumption The 15 pretraining datasets adequately cover the radio signal distribution for the downstream tasks.
    Stated in Section IV.A.1 without evidence of distribution coverage.
  • domain assumption Downstream evaluation splits are disjoint from pretraining data.
    Never stated; contradicted by the Table I and Table II overlap. This is the load-bearing premise for the transfer claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Radio-FM: A Foundation Model for Radio Signal Representation Learning and Its Applications." pith.science (2026). https://pith.science/paper/P4IA3FKL

@misc{pith2026260805793,
  author       = {Pith},
  title        = {Pith review of: Radio-FM: A Foundation Model for Radio Signal Representation Learning and Its Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P4IA3FKL}},
  note         = {Machine review of arXiv:2608.05793}
}
read the original abstract

Applying foundation models to the radio frequency (RF) domain presents unique challenges due to the intrinsic physical complexity of raw I/Q signals and the extreme heterogeneity of spectral data. In this paper, we present Radio-FM, a scalable family of foundation models designed for universal radio signal representation learning. Unlike standard architectures, Radio-FM employs dual-channel processing specifically optimized for I/Q independence while capturing cross-channel interactions through a lightweight attention mechanism. To scale pretraining across heterogeneous multi-source corpora with highly variable sequence lengths, we propose a token-budgeted dynamic batching strategy coupled with channel-independent masked reconstruction. We pretrain Radio-FM on a diverse collection of 15 datasets spanning modulation, radar, and communication domains, and rigorously evaluate it on 15 downstream benchmarks. Experimental results show Radio-FM achieves state-of-the-art performance on 13 of 15 benchmarks, consistently improving across modulation, radar, emitter identification, wireless technology recognition, and wireless interference identification. Notably, it exhibits superior few-shot transferability, significantly outperforming existing baselines in data-scarce regimes, validating its potential as a general-purpose backbone for radio signal understanding.

Figures

Figures reproduced from arXiv: 2608.05793 by the authors.

Figure 1
Figure 1. A representative electromagnetic ecosystem with het [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall training pipeline of Radio-FM, comprising three phases: (1) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Model architecture of Radio-FM. The encoder performs intra-channel self-attention and inter-channel interaction, and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Dataset collection of Radio-FM. TABLE II: Summary of downstream task datasets Dataset Name Task Type Signal Length No. of Classes RML2016.10a [15] AMR 128 11 RML2016.10b [15] AMR 128 10 RML2022 [43] AMR 128 11 RML2018.01a [40] AMR 1024 24 Sig2019-12 [17] AMR 512 12 HKD…
Figure 5
Figure 5. Figure 5: t-SNE visualization on RML2016.10a (AMR task) at 10dB SNR. From left to right: SpectrumFM, EMind, Radio-FM [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: t-SNE visualization on RadChar-Base (RWC task) at 10dB SNR. From left to right: SpectrumFM, EMind, Radio-FM [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: t-SNE visualization on TechRec (WTR task) at 10dB SNR. From left to right: SpectrumFM, EMind, Radio-FM Base, [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Pretraining Loss Convergence of Radio-FM Variants [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Ablation Studies on Radio-FM Scaling and Design Choices. (a) Average accuracy across model scales under different [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 25 canonical work pages

  1. [1]

    Building 6g radio foundation models with transformer architectures,

    A. Aboulfotouh, A. Eshaghbeigi, and H. Abou-Zeid, “Building 6g radio foundation models with transformer architectures,”arXiv preprint arXiv:2411.09996, 2024

    arXiv 2024

  2. [2]

    6g wavesfm: A foundation model for sensing, communication, and localization,

    A. Aboulfotouh, E. Mohammed, and H. Abou-Zeid, “6g wavesfm: A foundation model for sensing, communication, and localization,”arXiv preprint arXiv:2504.14100, 2025

    arXiv 2025

  3. [3]

    Towards channel foundation models (cfms): Motivations, methodologies and opportunities,

    J. Jiang, Y . Gao, X. Wu, and S. Xu, “Towards channel foundation models (cfms): Motivations, methodologies and opportunities,”arXiv preprint arXiv:2507.13637, 2025

    arXiv 2025

  4. [4]

    Wifo: Wireless foundation model for channel prediction,

    B. Liu, S. Gao, X. Liu, X. Cheng, and L. Yang, “Wifo: Wireless foundation model for channel prediction,”Science China Information Sciences, vol. 68, no. 6, p. 162302, 2025

    2025

  5. [5]

    Tiny federated wireless foundation models for resource constrained devices,

    M. Hallaq, F. M. A. Khan, A. Aboulfotouh, S. A. Hassan, K. Dev, M. T. Quasim, and H. Abou-Zeid, “Tiny federated wireless foundation models for resource constrained devices,”IEEE Internet of Things Journal, 2025

    2025

  6. [6]

    Scale what counts, mask what matters: Evaluating foundation models for zero-shot cross- domain wi-fi sensing,

    C. Jiang, Y . Yan, Y . Wang, C. T. Chou, and W. Hu, “Scale what counts, mask what matters: Evaluating foundation models for zero-shot cross- domain wi-fi sensing,”arXiv preprint arXiv:2511.18792, 2025

    arXiv 2025

  7. [7]

    Multimodal wireless foundation models,

    A. Aboulfotouh and H. Abou-Zeid, “Multimodal wireless foundation models,”arXiv preprint arXiv:2511.15162, 2025

    arXiv 2025

  8. [8]

    Large wireless localization model (lwlm): A foundation model for positioning in 6g networks,

    G. Pan, K. Huang, H. Chen, S. Zhang, C. H ¨ager, and H. Wymeersch, “Large wireless localization model (lwlm): A foundation model for positioning in 6g networks,”arXiv preprint arXiv:2505.10134, 2025

    arXiv 2025

Show all 43 references
  1. [9]

    Rf-diffusion: Radio signal generation via time-frequency diffusion,

    G. Chi, Z. Yang, C. Wu, J. Xu, Y . Gao, Y . Liu, and T. X. Han, “Rf-diffusion: Radio signal generation via time-frequency diffusion,” in Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pp. 77–92, 2024

    2024

  2. [10]

    Wirelessgpt: A generative pre-trained multi-task learning framework for wireless communication,

    T. Yang, P. Zhang, M. Zheng, Y . Shi, L. Jing, J. Huang, and N. Li, “Wirelessgpt: A generative pre-trained multi-task learning framework for wireless communication,”arXiv preprint arXiv:2502.06877, 2025

    2025 arXiv

  3. [11]

    Ris-mae: A self-supervised modulation classification method based on raw iq signals and masked autoencoder,

    Y . Liu, M. Liu, W. Xie, X. Liu, W. Liu, Y . Sun, X. Qiu, C. Yuan, and J. Li, “Ris-mae: A self-supervised modulation classification method based on raw iq signals and masked autoencoder,”arXiv preprint arXiv:2508.00274, 2025

    2025 arXiv

  4. [12]

    Spectrumfm: A foundation model for intelligent spectrum management,

    F. Zhou, C. Liu, H. Zhang, W. Wu, Q. Wu, T. Q. S. Quek, and C.- B. Chae, “Spectrumfm: A foundation model for intelligent spectrum management,”IEEE Journal on Selected Areas in Communications, pp. 1–1, 2025

    2025

  5. [13]

    Emind: A foundation model for multi-task elec- tromagnetic signals understanding,

    L. Luo, W. Gui, Y . Liu, Z. Zhang, Y . Zhang, F. Wang, Z. Guo, Z. Ma, X. Liu, H. He,et al., “Emind: A foundation model for multi-task elec- tromagnetic signals understanding,”arXiv preprint arXiv:2508.18785, 2025

    2025 arXiv

  6. [14]

    Large- scale real-world radio signal recognition with deep learning,

    Y . Tu, Y . Lin, H. Zha, J. Zhang, Y . Wang, G. GUI, and S. Mao, “Large- scale real-world radio signal recognition with deep learning,”Chinese Journal of Aeronautics, vol. 35, no. 9, pp. 35–48, 2022

    2022

  7. [15]

    Convolutional radio modula- tion recognition networks,

    T. J. O’Shea, J. Corgan, and T. C. Clancy, “Convolutional radio modula- tion recognition networks,” inInternational conference on engineering applications of neural networks, pp. 213–226, Springer, 2016

    2016

  8. [16]

    Robust and fast automatic modulation classification with cnn under multipath fading channels,

    K. Tekbıyık, A. R. Ekti, A. G ¨orc ¸in, G. K. Kurt, and C. Kec ¸eci, “Robust and fast automatic modulation classification with cnn under multipath fading channels,” in2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring), pp. 1–6, 2020. 12 Tiny 2.38M Small 6.79M Base...

    2020

  9. [17]

    Signet: A novel deep learning framework for radio signal classification,

    Z. Chen, H. Cui, J. Xiang, K. Qiu, L. Huang, S. Zheng, S. Chen, Q. Xuan, and X. Yang, “Signet: A novel deep learning framework for radio signal classification,”IEEE Transactions on Cognitive Communi- cations and Networking, vol. 8, no. 2, pp. 529–541, 2021

    2021

  10. [18]

    Contour stella image and deep learning for signal recognition in the physical layer,

    Y . Lin, Y . Tu, Z. Dou, L. Chen, and S. Mao, “Contour stella image and deep learning for signal recognition in the physical layer,”IEEE Transactions on Cognitive Communications and Networking, vol. 7, no. 1, pp. 34–46, 2020

    2020

  11. [19]

    Complex-valued networks for automatic modulation classification,

    Y . Tu, Y . Lin, C. Hou, and S. Mao, “Complex-valued networks for automatic modulation classification,”IEEE Transactions on Vehicular Technology, vol. 69, no. 9, pp. 10085–10089, 2020

    2020

  12. [20]

    Semi-supervised learning with generative adversarial networks on digital signal modulation classifica- tion.,

    Y . Tu, Y . Lin, J. Wang, and J.-U. Kim, “Semi-supervised learning with generative adversarial networks on digital signal modulation classifica- tion.,”Computers, Materials & Continua, vol. 55, no. 2, 2018

    2018

  13. [21]

    Avgnet: Adaptive visibility graph neural network and its application in modulation classification,

    Q. Xuan, J. Zhou, K. Qiu, Z. Chen, D. Xu, S. Zheng, and X. Yang, “Avgnet: Adaptive visibility graph neural network and its application in modulation classification,”IEEE Transactions on Network Science and Engineering, vol. 9, no. 3, pp. 1516–1526, 2022

    2022

  14. [22]

    Dtsg-net: Dynamic time series graph neural network and it’s application in modulation recognition,

    P. Yin, J. Zhou, Y . Ge, and Z. Chen, “Dtsg-net: Dynamic time series graph neural network and it’s application in modulation recognition,” IEEE Internet of Things Journal, 2024

    2024

  15. [23]

    Lstm framework for classification of radar and communications signals,

    V . Clerico, J. Gonz´alez-L´opez, G. Agam, and J. Grajal, “Lstm framework for classification of radar and communications signals,” in2023 IEEE Radar Conference (RadarConf23), pp. 1–6, 2023

    2023

  16. [24]

    Multi- task learning for radar signal characterisation,

    Z. Huang, A. Pemasiri, S. Denman, C. Fookes, and T. Martin, “Multi- task learning for radar signal characterisation,” in2023 IEEE In- ternational Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pp. 1–5, IEEE, 2023

    2023

  17. [25]

    Wisig: A large-scale wifi signal dataset for receiver and channel agnostic rf fingerprinting,

    S. Hanna, S. Karunaratne, and D. Cabric, “Wisig: A large-scale wifi signal dataset for receiver and channel agnostic rf fingerprinting,”IEEE Access, vol. 10, pp. 22808–22818, 2022

    2022

  18. [26]

    Trust in 5g open rans through machine learning: Rf fingerprinting on the powder pawr platform,

    G. Reus-Muns, D. Jaisinghani, K. Sankhe, and K. R. Chowdhury, “Trust in 5g open rans through machine learning: Rf fingerprinting on the powder pawr platform,” inGLOBECOM 2020-2020 IEEE Global Communications Conference, pp. 1–6, IEEE, 2020

    2020

  19. [27]

    Radio frequency fingerprint identification towards statistical and deep learning features: Review, recent results and future directions,

    G. Yan, X. Fu, Y . Wang, Q. Zhang, and G. Gui, “Radio frequency fingerprint identification towards statistical and deep learning features: Review, recent results and future directions,”Peer-to-Peer Networking and Applications, vol. 18, no. 3, pp. 1–25, 2025

    2025

  20. [28]

    Tfmix: A robust time-frequency mixing approach for domain generalization in specific emitter identification,

    S. Wang, H. Wan, F. Wang, Y . Wang, Y . Lin, and G. Gui, “Tfmix: A robust time-frequency mixing approach for domain generalization in specific emitter identification,”IEEE Transactions on Cognitive Communications and Networking, 2025

    2025

  21. [29]

    Towards low-complexity wireless technology classification across multiple environments,

    J. Fontaine, E. Fonseca, A. Shahid, M. Kist, L. A. DaSilva, I. Moer- man, and E. De Poorter, “Towards low-complexity wireless technology classification across multiple environments,”Ad Hoc Networks, vol. 91, p. 101881, 2019

    2019

  22. [30]

    Multi-band sub-ghz technology recognition on nvidia’s jetson nano,

    J. Fontaine, A. Shahid, R. Elsas, A. Seferagic, I. Moerman, and E. De Poorter, “Multi-band sub-ghz technology recognition on nvidia’s jetson nano,” in2020 IEEE 92nd Vehicular Technology Conference (VTC2020-Fall), pp. 1–7, IEEE, 2020

    2020

  23. [31]

    Wireless interference iden- tification with convolutional neural networks,

    M. Schmidt, D. Block, and U. Meier, “Wireless interference iden- tification with convolutional neural networks,” in2017 IEEE 15th international conference on industrial informatics (INDIN), pp. 180– 185, IEEE, 2017

    2017

  24. [32]

    Deep learning for interference identification: Band, training snr, and sample selection,

    X. Zhang, T. Seyfi, S. Ju, S. Ramjee, A. El Gamal, and Y . C. Eldar, “Deep learning for interference identification: Band, training snr, and sample selection,” in2019 IEEE 20th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pp. 1–5, IEEE, 2019

    2019

  25. [33]

    Iqfm a wireless foundational model for i/q streams in ai-native 6g,

    O. Mashaal and H. Abou-Zeid, “Iqfm a wireless foundational model for i/q streams in ai-native 6g,”arXiv preprint arXiv:2506.06718, 2025

    2025 arXiv

  26. [34]

    A foundation model for wireless technology recognition and localiza- tion tasks,

    M. Cheraghinia, E. De Poorter, J. Fontaine, M. Debbah, and A. Shahid, “A foundation model for wireless technology recognition and localiza- tion tasks,”IEEE Open Journal of the Communications Society, vol. 6, pp. 9879–9896, 2025

    2025

  27. [35]

    Skyllm: Enabling trustworthy uav rf surveillance with foundation models for open-world signal recognition,

    W. Zhang, G. Gui, X. Wang, H. Sari, and F. Adachi, “Skyllm: Enabling trustworthy uav rf surveillance with foundation models for open-world signal recognition,”IEEE Wireless Communications, pp. 1–8, 2026

    2026

  28. [36]

    Roformer: En- hanced transformer with rotary position embedding,

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “Roformer: En- hanced transformer with rotary position embedding,”Neurocomputing, vol. 568, p. 127063, 2024

    2024

  29. [37]

    Go- ing deeper with image transformers,

    H. Touvron, M. Cord, A. Sablayrolles, G. Synnaeve, and H. J ´egou, “Go- ing deeper with image transformers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 32–42, 2021

    2021

  30. [38]

    Deep networks with stochastic depth,

    G. Huang, Y . Sun, Z. Liu, D. Sedra, and K. Q. Weinberger, “Deep networks with stochastic depth,” inEuropean conference on computer vision, pp. 646–661, Springer, 2016

    2016

  31. [39]

    Toward next-generation signal intelligence: A hybrid knowledge and data-driven deep learning framework for radio signal classification,

    S. Zheng, X. Zhou, L. Zhang, P. Qi, K. Qiu, J. Zhu, and X. Yang, “Toward next-generation signal intelligence: A hybrid knowledge and data-driven deep learning framework for radio signal classification,” IEEE Transactions on Cognitive Communications and Networking, vol. 9, no. ...

    2023

  32. [40]

    Over-the-air deep learning based radio signal classification,

    T. J. O’Shea, T. Roy, and T. C. Clancy, “Over-the-air deep learning based radio signal classification,”IEEE Journal of Selected Topics in Signal Processing, vol. 12, no. 1, pp. 168–179, 2018

    2018

  33. [41]

    Dataset for modulation classification and signal type classification for multi-task and single task learning,

    A. Jagannath and J. Jagannath, “Dataset for modulation classification and signal type classification for multi-task and single task learning,” Computer Networks, vol. 199, p. 108441, 2021

    2021

  34. [42]

    Large scale radio frequency signal classification,

    L. Boegner, M. Gulati, G. Vanhoy, P. Vallance, B. Comar, S. Kokalj- Filipovic, C. Lennon, and R. D. Miller, “Large scale radio frequency signal classification,” 2022. arXiv:2207.09918

    2022 arXiv

  35. [43]

    Rml22: Realistic dataset generation for wireless modulation classification,

    V . Sathyanarayanan, P. Gerstoft, and A. E. Gamal, “Rml22: Realistic dataset generation for wireless modulation classification,”IEEE Trans- actions on Wireless Communications, vol. 22, no. 11, pp. 7663–7675, 2023

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.