Pith. sign in

REVIEW 3 major objections 7 minor 67 references

A single heterogeneous edge SoC can run three RF intelligence tasks end-to-end in 98 µs without preprocessing or hardware redesign.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 08:18 UTC pith:FQC5MVO5

load-bearing objection Solid multi-task RF edge accelerator with real board numbers; the latency claim is fine, the “keep pace with 61.44 MSPS” framing is not. the 3 major comments →

arxiv 2607.24669 v1 pith:FQC5MVO5 submitted 2026-07-27 cs.AR

A Heterogeneous Neural Network Accelerator for End-to-End Multitask RF Signal Recognition

Zhifan Song , Haralampos-G. Stratigopoulos , Hassan Aboushady This is my paper
classification cs.AR
keywords AI hardware accelerationheterogeneous computingRF modulation recognitionFPGA SoClearnable decimatorcovert channel detectionGNSS jammingedge inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Modern radios need to classify what is on the air—modulation type, covert channels, and jamming—in real time at the device, not in the cloud. This paper shows that one compact attention-enhanced CNN, fed by a learnable streaming decimator that works directly on ADC samples, can handle three such tasks when the heavy layers run on a small FPGA accelerator and the flexible layers run on the onboard CPU. Cache-aligned co-scheduling keeps the two sides coherent and balanced. On three public datasets the system reaches at least 99% accuracy above 4 dB SNR for modulation recognition, 90% for hardware-Trojan covert-channel detection, and 99.5% for GNSS jamming, all at 98 µs per frame and roughly half a watt of programmable-logic power. The practical claim is that multi-task spectrum intelligence becomes feasible on ordinary embedded SoCs without redesigning silicon for each new RF problem.

Core claim

A heterogeneous CPU–accelerator SoC that pairs a learnable streaming decimator (LSDec) with a dual-pipeline fused convolution–max-pool engine and cache-line-aligned co-scheduling can deliver end-to-end multi-task RF signal recognition—AMR, HT-CC detection, and GNSS jamming classification—directly from raw I/Q streams at 98 µs per frame, sub-watt PL power, and high accuracy, while allowing task changes by loading new weights and CPU-side hyperparameters only.

What carries the argument

LSDec plus dual-pipeline fused Conv–MaxPool with cache-aligned heterogeneous co-scheduling (Algorithm 1): LSDec performs on-the-fly learnable decimation at the ADC interface; the fused engine halves memory traffic; the scheduler partitions filters so CPU and accelerator finish together without cache-line conflicts.

Load-bearing premise

That a single Cortex-A53 core plus strict cache-line-aligned filter splits will always keep the shared-memory pipeline correct and inside 98 µs when the radio front-end, CPU load, or a new RF task differs from the three evaluated cases.

What would settle it

Measure end-to-end latency and classification accuracy on the same three datasets while forcing multi-core contention or a fourth RF task whose hyperparameters cannot be absorbed by the CPU-side layers; any systematic breach of 98 µs or large accuracy drop falsifies the real-time multi-task claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Embedded SDRs and UAVs can run modulation recognition, covert-channel detection, and jamming classification on one low-power SoC without cloud offload.
  • New RF tasks can be added by retraining weights and tuning CPU-executed layers rather than resynthesizing the accelerator.
  • Direct ADC-to-AI streaming removes full-frame buffering, so front-end sample rates no longer force frame drops on edge devices.
  • Sub-watt PL power and ~20 k parameters make the same architecture a candidate for ASIC or lower-end FPGA spectrum monitors.

Where Pith is reading between the lines

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

  • The same co-scheduling pattern could be reused for other streaming sensor modalities (audio, radar, vibration) that also face ADC-to-edge speed mismatch.
  • If multi-core RAP is left always-on, dynamic power of the accelerator should fall roughly in proportion to idle processing elements, suggesting an automatic energy–throughput knob.
  • Task multiplexing at 98 µs opens the possibility of frame-by-frame switching among spectrum, security, and navigation classifiers on a single radio link.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The manuscript presents a heterogeneous SoC accelerator for multi-task RF signal recognition on a Zynq UltraScale+ ZCU104. A compact attention-enhanced CNN with a learnable streaming decimator (LSDec, k=s=D) is partitioned across FPGA fabric (dual-pipeline fused Conv–MaxPool engine, folded PWConv for dense layers) and NEON-optimized Cortex-A53 kernels, with a cache-line-aligned co-scheduling algorithm (Alg. 1) governing the filter split. The same topology serves three tasks — AMR on RadioML2018 (≥99% above 4 dB SNR, 73.3% all-SNR average float / 72.2% quantized), HT-CC detection (90.1% tuned / 89.3% quantized), and GNSS jamming (99.5% float / 98.5% quantized) — with task switching by weight/hyperparameter reload only. Board-measured end-to-end latency is 98 µs per 1024-sample frame including synchronization, cache flush, and DMA, at 0.54 W PL power, 38k LUTs / 26k FFs / 405 DSPs. An ablation (Table VI) attributes gains to LSDec (+1.8% over baseline while cutting kMACs 88%), attention (+1.4%), and filter reduction (−0.2% accuracy for ~45% parameter savings).

Significance. If the throughput clarification lands acceptably, this is a solid and useful contribution: to the reviewer's knowledge the first heterogeneous CPU+FPGA accelerator demonstrated across three distinct RF intelligence tasks (AMR, covert-channel detection, GNSS jamming) on a single low-cost SoC. The strengths are concrete and falsifiable: a learnable streaming decimator with a clean stationary-kernel hardware realization (one MAC per channel, no shift registers), a component-level ablation isolating LSDec/attention/filter-reduction contributions (Table VI), an explicit cache-line-alignment co-scheduling algorithm with a worked numerical case study, board-measured latency including sync/flush/DMA overheads, and same-split GPU baselines for ten comparison models. The W8A16-INT quantization with shift-only requantization, and the PWConv folding that eliminates a dedicated dense-layer IP, are pragmatic engineering results of direct interest to the embedded-SDR community.

major comments (3)
  1. [§IV-A, §VI-A, §VII-B] The manuscript motivates the streaming design by criticizing buffering-based approaches for 'unavoidable latency and further frame drops' at ADC rates 'up to 61.44 MSPS' (§IV-A), and claims raw I/Q streams are fed 'directly' for 'real-time' inference (§I). However, the measured 98 µs per-frame latency caps sustained processing at ~10.2 kFPS, i.e. ~10.4 MSPS effective for 1024-sample frames. At the cited 61.44 MSPS a frame arrives every ~16.7 µs, so unless the layer pipeline overlaps across frames (never stated), only ~1 in 6 frames can be serviced. The paper never specifies the ADC sample rate or duty cycle at which the 98 µs figure was measured, nor the frame-drop behavior of the deployed system. This is load-bearing in two places: (i) the 'keep pace with RF front-end ADC output rates' framing (§II-A1, §IV-A), and (ii) the two security tasks, where HT-CC events and pulsed GNSS jammers a
  2. [§VII-B, Table VII, Abstract] The abstract and Conclusion headline '0.54 W' PL power, but the design's central idea is to offload work to the Cortex-A53 (attention and dense layers run entirely on the CPU; §IV-C, §VII-A), and §VII-B itself cites 2.64 W dynamic power for the A53 cluster. For a paper whose stated contribution is power-efficient edge inference, the system-level power for the 98 µs configuration (PL + the CPU cores actually used) should be reported alongside the PL-only figure, and Table VII's comparison with sub-watt PL implementations should note that those baselines do not rely on a multi-watt CPU in the inference path. The Fig. 9 caption's assumption that 'an embedded CPU is always present' covers control overhead, not CPU-executed network layers; the two should be distinguished.
  3. [Abstract, §VII-D] The abstract reports 99.5% on GNSS jamming and 90% on HT-CC, but the deployed quantized model achieves 98.5% and 89.3% respectively (§VII-D); the abstract numbers are the floating-point GPU results. Since the paper's claim is about a hardware system, the abstract should either report the on-hardware quantized accuracies or explicitly label which figures are float vs. deployed. The same applies to the AMR claim: '≥99% above 4 dB' holds for both models per Fig. 5, but the 73.3% all-SNR average quoted in §VI-A1 is the float figure (72.2% quantized).
minor comments (7)
  1. [§VII-C] The case study states '7986 mod 64 = 34'; the correct value is 50 (7986 = 124×64 + 50). This does not affect the conclusion — 32 filters × 242 B = 7744 B is indeed cache-line aligned (7744 = 121×64) — but the intermediate arithmetic should be corrected.
  2. [§VI-A3] 'five to six significantly lower computational complexity' is garbled; presumably 'five to six orders of magnitude lower' is intended. Note that relative to some baselines (e.g. CGDNet at 1,819,486 kMACs vs. 374) the gap is closer to 3.7 orders; please align the wording with the actual range in Table V.
  3. [§V-A] The sample rate and capture duration of each dataset (RadioML2018, HT-CC at bladeRF rate, GNSS Zenodo set) are not stated in §V-A. These are needed to interpret the per-frame latency against real stream rates, especially given Major Comment 1.
  4. [Table V] The GNSS dataset SNR is described as 'one fixed SNR' but the value is never given; please state it in Table V's caption or §V-A3.
  5. [§IV-D, Algorithm 1] In Alg. 1 the loop variable K (line 7) shadows the filter-count symbol K_a used in the surrounding text, and the unconstrained optimizer on line 5 uses t_a/t_c while the case study quotes total layer times T_a = 42 µs, T_c = 353 µs; a sentence clarifying per-channel vs. total latency would prevent confusion (the numerical outcome, ~32–33 filters, is consistent either way).
  6. [§V-B] No code, model weights, or bitstream availability is mentioned. Given the paper's emphasis on reproducibility (identical splits, same-GPU baselines), a release of the training/quantization scripts and C/Neon kernels would substantially strengthen the work.
  7. [Figs. 9–11] Several figures are dense at column width: Fig. 10's log-scale latency bars would benefit from numeric labels (the 1 µs dense layer is 'discarded' from the figure but referenced in the text), and Fig. 11's two confusion matrices should state class labels or a legend for the six GNSS classes.

Circularity Check

0 steps flagged

Empirical hardware/ML systems paper: accuracy, latency, and power are measured against external datasets and board profiling, not derived by construction from fitted identities or self-citation chains.

full rationale

The paper’s load-bearing claims are empirical: classification accuracy on RadioML2018, HOST HT-CC, and GNSS-jamming datasets; end-to-end latency (98 µs) and PL power/resources on ZCU104; and a cache-aligned CPU–accelerator partition chosen from profiled per-channel latencies. None of these reduce to their inputs by definition. The co-scheduling rule T(K_a)=max(K_a t_a,(N−K_a)t_c) with cache-line alignment is standard load balancing applied to measured t_a/t_c, not a tautology. Self-citations to the authors’ prior FPGA accelerator [19] and HT-CC dataset [5] supply background and a training corpus; they do not underwrite a uniqueness theorem or force the reported accuracy/latency numbers. Model hyperparameters are trained end-to-end and quantized; reported accuracies are held-out test metrics, not refits of the same quantities. No self-definitional loop, fitted-input-as-prediction, or ansatz-smuggled-via-citation pattern appears in the derivation chain. Throughput-vs-latency framing gaps (if any) are claim-scope issues, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Load-bearing content is engineering design choices and empirical fits (architecture hyperparameters, decimation factor, quantization widths, partition Ka), plus standard assumptions about ARM cache coherency and public RF datasets representing deployment. No new physical entities; ‘invented’ items are named modules/algorithms.

free parameters (5)
  • LSDec decimation factor D (k=s=D) = D=8 (AMR example)
    Chosen so sequence length and MAC count drop (e.g. 1024→128); drives accuracy–cost tradeoff and streaming schedule.
  • Conv filter counts and dense widths (e.g. 36 filters, 32/48 neurons) = 36 filters; HT-CC tuned 29k params
    Hand/task-tuned relative to baseline CNN; task-tuned HT-CC variant widens attention bottleneck and dense layer.
  • Attention reduction ratio = task-dependent (e.g. 9 or 3)
    Task-configurable SE-style bottleneck (e.g. 36→4 default; HT-CC 9→3) selected for accuracy on each dataset.
  • W8A16-INT quantization scales (right-shift requantization) = INT8 weights, INT16 activations
    Bit-widths and scales set post-training to preserve RF dynamic range; accuracy after quant is part of the claimed operating point.
  • Co-schedule partition Ka (cache-line aligned) = e.g. Ka=32 of N=36
    Chosen from profiled ta, tc and 64-byte alignment; determines reported latency and resource split (e.g. 32/4 on Conv2d).
axioms (5)
  • domain assumption Public dataset frames (RadioML2018, HT-CC, GNSS jamming) and fixed splits are adequate proxies for embedded real-time RF workloads.
    All accuracy claims rest on these offline sets; closed-loop ADC/SDR channel effects are not re-validated end-to-end in the results.
  • domain assumption ARM Cortex-A53 64-byte cache lines and clean-invalidate over CPU-written regions suffice for safe shared-DDR co-execution when Ka*B is line-aligned.
    Correctness of heterogeneous writes (§IV-D, Alg. 1) depends on this memory model.
  • ad hoc to paper Same CNN topology with only weight/hyperparameter changes is sufficient across AMR, HT-CC, and GNSS tasks without FPGA resynthesis.
    Central flexibility claim; supported on three tasks but assumed to generalize.
  • ad hoc to paper Fused Conv–MaxPool with fixed stride-2 max-pool after the second conv is an acceptable specialization for all target tasks.
    Stated tradeoff in §VII-B; underpins dual-pipeline resource and latency numbers.
  • standard math Standard CNN/MAC complexity and INT quantization arithmetic behave as modeled in PyTorch then C/Verilog rewrite.
    Training and kMAC counts rely on ordinary DL numerics.
invented entities (3)
  • LSDec (learnable streaming decimator) independent evidence
    purpose: On-the-fly ADC-rate decimation and feature extraction with stationary shared I/Q Conv1d to avoid full-frame buffering.
    Named module central to end-to-end streaming claim; implemented as constrained Conv1d, not a new physical law.
  • Dual-pipeline fused Conv–MaxPool engine independent evidence
    purpose: Compute two spatial outputs and pool on the fly to cut DDR traffic and match Conv2d latency.
    Hardware organization claimed novel in this SoC; evidence is synthesis and latency on ZCU104.
  • Cache-aligned heterogeneous co-scheduling algorithm (Alg. 1) independent evidence
    purpose: Choose Ka minimizing max(Ka ta,(N-Ka)tc) subject to cache-line alignment and optional multi-core symmetry.
    Scheduling rule load-bearing for deterministic shared-memory execution; validated by Conv2d case study.

pith-pipeline@v1.2.0-grok45-kimik3 · 26683 in / 3813 out tokens · 75698 ms · 2026-07-31T08:18:48.979197+00:00 · methodology

0 comments
read the original abstract

This paper presents a heterogeneous neural network accelerator for multi-task RF signal recognition, supporting automatic modulation recognition (AMR), hardware-Trojan covert channel (HT-CC) detection, and GNSS jamming classification. We introduce a compact attention-enhanced convolutional neural network (CNN) combined with LSDec, a learnable streaming decimator that enables adaptive temporal downsampling and flexible input lengths. The hardware architecture integrates a novel dual-pipeline, fused convolution-pooling engine with DMA-based streaming to minimize memory traffic and latency. Co-execution scheduling on the accelerator and SIMD-optimized CPU kernels reduces hardware resource usage while preserving high performance and task-level flexibility. Across three datasets, the proposed system achieves $\geq$ 99% average accuracy above 4 dB Signal-to-Noise Ratios (SNRs) on the RadioML2018 dataset for AMR, 90% on the HT-CC dataset, and 99.5% on the GNSS-Jamming dataset. The accelerator sustains an end-to-end inference latency of 98 $\mu$s per frame, demonstrating its effectiveness for low-power, latency-critical multi-task spectrum-intelligence applications on embedded and edge devices.

Figures

Figures reproduced from arXiv: 2607.24669 by Haralampos-G. Stratigopoulos, Hassan Aboushady, Zhifan Song.

Figure 1
Figure 1. Figure 1: End-to-end RF spectrum monitoring system with dedicated AI SoC. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the proposed integer-quantized attention-infused CNN model for three RF tasks, namely AMR, HT-CC detection, and GNSS jamming [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Microarchitecture of the proposed heterogeneous AI accelerator SoC for multi-task RF signal recognition. It integrates a quad-core ARM Cortex-A53 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hardware architecture of LSDec. Utilizing a shared, stationary weights [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Figure of merit: accuracy–complexity trade-off for AMR measured as [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: The confusion matrix at SNR = 0 dB for the AMR task using [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Accuracy vs. SNR curve on the HT-CC dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FPGA resource and power comparison at 125 MHz. Scenario 1: [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Confusion matrix on GNSS jamming and HT-CC (SNR=13 dB) [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗

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

67 extracted references · 1 linked inside Pith

  1. [1]

    Redefining Wireless Communication for 6G: Signal Processing Meets Deep Learning With Deep Unfolding,

    A. Jagannath, J. Jagannath, and T. Melodia, “Redefining Wireless Communication for 6G: Signal Processing Meets Deep Learning With Deep Unfolding,”IEEE Transactions on Artificial Intelligence, vol. 2, no. 6, pp. 528–536, 2021

    2021

  2. [2]

    AI Empowered Wireless Communications: From Bits to Semantics,

    Z. Qin, L. Liang, Z. Wang, S. Jin, X. Tao, W. Tong, and G. Y . Li, “AI Empowered Wireless Communications: From Bits to Semantics,” Proceedings of the IEEE, vol. 112, no. 7, pp. 621–652, 2024

    2024

  3. [3]

    AI-Managed Cognitive Radio Digitizers,

    J. M. de la Rosa, “AI-Managed Cognitive Radio Digitizers,”IEEE Circuits and Systems Magazine, vol. 22, no. 1, pp. 10–39, 2022

    2022

  4. [4]

    A Review of Research on Signal Modulation Recognition Based on Deep Learning,

    W. Xiao, Z. Luo, and Q. Hu, “A Review of Research on Signal Modulation Recognition Based on Deep Learning,”Electronics, vol. 11, no. 17, 2022

    2022

  5. [5]

    Covert Communication Channels Based on Hardware Trojans: Open- Source Dataset and AI-Based Detection,

    A. R. D ´ıaz-Rizo, A. Emad, H. Aboushady, and H.-G. Stratigopoulos, “Covert Communication Channels Based on Hardware Trojans: Open- Source Dataset and AI-Based Detection,” in2024 IEEE International Symposium on Hardware Oriented Security and Trust (HOST), pp. 101– 106, 2024

    2024

  6. [6]

    Jammer Classi- fication in GNSS Bands Via Machine Learning Algorithms,

    R. Morales Ferre, A. de la Fuente, and E. S. Lohan, “Jammer Classi- fication in GNSS Bands Via Machine Learning Algorithms,”Sensors, vol. 19, no. 22, 2019

    2019

  7. [7]

    Study of modulation recogni- tion based on HOCs and SVM,

    H. Gang, L. Jiandong, and L. Donghua, “Study of modulation recogni- tion based on HOCs and SVM,” in2004 IEEE 59th Vehicular Technology Conference. VTC 2004-Spring (IEEE Cat. No.04CH37514), vol. 2, pp. 898–902 V ol.2, 2004

    2004

  8. [8]

    Deep Learning Models for Wireless Signal Classification With Dis- tributed Low-Cost Spectrum Sensors,

    S. Rajendran, W. Meert, D. Giustiniano, V . Lenders, and S. Pollin, “Deep Learning Models for Wireless Signal Classification With Dis- tributed Low-Cost Spectrum Sensors,”IEEE Transactions on Cognitive Communications and Networking, vol. 4, no. 3, pp. 433–445, 2018

    2018

  9. [9]

    A Spatiotemporal Multi-Channel Learning Framework for Automatic Modulation Recognition,

    J. Xu, C. Luo, G. Parr, and Y . Luo, “A Spatiotemporal Multi-Channel Learning Framework for Automatic Modulation Recognition,”IEEE Wireless Communications Letters, vol. 9, no. 10, pp. 1629–1632, 2020

    2020

  10. [10]

    CNN-Based Automatic Modulation Classification for Beyond 5G Communications,

    A. P. Hermawan, R. R. Ginanjar, D.-S. Kim, and J.-M. Lee, “CNN-Based Automatic Modulation Classification for Beyond 5G Communications,” IEEE Communications Letters, vol. 24, no. 5, pp. 1038–1041, 2020

    2020

  11. [11]

    CGDNet: Effi- cient Hybrid Deep Learning Model for Robust Automatic Modulation Recognition,

    J. N. Njoku, M. E. Morocho-Cayamcela, and W. Lim, “CGDNet: Effi- cient Hybrid Deep Learning Model for Robust Automatic Modulation Recognition,”IEEE Networking Letters, vol. 3, no. 2, pp. 47–51, 2021

    2021

  12. [12]

    End-to-End Learning From Spectrum Data: A Deep Learning Approach for Wire- less 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 Wire- less Signal Identification in Spectrum Monitoring Applications,”IEEE Access, vol. 6, pp. 18484–18501, 2018

    2018

  13. [13]

    Big Spectrum Data and Deep Learning Techniques for Cognitive Wireless Networks,

    P. D. Choudhury, A. Bora, and K. K. Sarma, “Big Spectrum Data and Deep Learning Techniques for Cognitive Wireless Networks,” inDeep Learning and Neural Networks: Concepts, Methodologies, Tools, and Applications(I. R. M. Association, ed.), pp. 994–1015, Hershey, PA, USA: IGI Global, 2020

    2020

  14. [14]

    An End-to-End Neuromorphic Radio Classification System With an Efficient Sigma-Delta-Based Spike Encoding Scheme,

    W. Guo, K. Yang, H.-G. Stratigopoulos, H. Aboushady, and K. N. Salama, “An End-to-End Neuromorphic Radio Classification System With an Efficient Sigma-Delta-Based Spike Encoding Scheme,”IEEE Transactions on Artificial Intelligence, vol. 5, no. 4, pp. 1869–1881, 2024

    2024

  15. [15]

    An On-chip Accelerator with Hybrid Machine Learning for Modulation Classification of Radio Frequency Signals,

    K. Jung, J. Woo, and S. Mukhopadhyay, “An On-chip Accelerator with Hybrid Machine Learning for Modulation Classification of Radio Frequency Signals,” in2022 IEEE/MTT-S International Microwave Symposium - IMS 2022, pp. 487–490, 2022

    2022

  16. [16]

    Real-Time and Embedded Deep Learning on FPGA for RF Signal Classification,

    S. Soltani, Y . E. Sagduyu, R. Hasan, K. Davaslioglu, H. Deng, and T. Erpek, “Real-Time and Embedded Deep Learning on FPGA for RF Signal Classification,” inMILCOM 2019 - 2019 IEEE Military Communications Conference (MILCOM), pp. 1–6, 2019

    2019

  17. [17]

    On-Chip Acceleration of RF Signal Modulation Classification With Short-Time Fourier Transform and Convolutional Neural Network,

    K. Jung, J. Woo, and S. Mukhopadhyay, “On-Chip Acceleration of RF Signal Modulation Classification With Short-Time Fourier Transform and Convolutional Neural Network,”IEEE Access, vol. 11, pp. 144051– 144063, 2023

    2023

  18. [18]

    Deep Learning Modulation Recognition for RF Spec- trum Monitoring,

    A. Emad, H. Mohamed, A. Farid, M. Hassan, R. Sayed, H. Aboushady, and H. Mostafa, “Deep Learning Modulation Recognition for RF Spec- trum Monitoring,” in2021 IEEE International Symposium on Circuits and Systems (ISCAS), pp. 1–5, 2021

    2021

  19. [19]

    A Power-Efficient Attention-Infused CNN Hardware Accelerator for RF Spectrum Monitoring,

    Z. Song, A. E. Abdelazim, P. Bazargan Sabet, F. Wajsb ¨urt, H.-G. Stratigopoulos, and H. Aboushady, “A Power-Efficient Attention-Infused CNN Hardware Accelerator for RF Spectrum Monitoring,” in2025 IEEE International Symposium on Circuits and Systems (ISCAS), pp. 1–5, 2025

    2025

  20. [20]

    Hetero2Pipe: Pipelining Multi-DNN Inference on Heterogeneous Mo- bile Processors under Co-Execution Slowdown,

    Y . Shen, Z. Wang, T. Wang, C. Gu, Z. Wen, Y . Shu, and C. Wang, “Hetero2Pipe: Pipelining Multi-DNN Inference on Heterogeneous Mo- bile Processors under Co-Execution Slowdown,” in2025 IEEE 45th International Conference on Distributed Computing Systems (ICDCS), pp. 483–493, 2025

    2025

  21. [21]

    Modulation Classification Based on Signal Constellation Diagrams and Deep Learning,

    S. Peng, H. Jiang, H. Wang, H. Alwageed, Y . Zhou, M. M. Sebdani, and Y .-D. Yao, “Modulation Classification Based on Signal Constellation Diagrams and Deep Learning,”IEEE Transactions on Neural Networks and Learning Systems, vol. 30, no. 3, pp. 718–727, 2019

    2019

  22. [22]

    Automatic modulation recognition based on CNN and GRU,

    F. Liu, Z. Zhang, and R. Zhou, “Automatic modulation recognition based on CNN and GRU,”Tsinghua Science and Technology, vol. 27, no. 2, pp. 422–431, 2022

    2022

  23. [23]

    Automatic Modulation Recognition Based on Adaptive Attention Mechanism and ResNeXt WSL Model,

    Z. Liang, M. Tao, L. Wang, J. Su, and X. Yang, “Automatic Modulation Recognition Based on Adaptive Attention Mechanism and ResNeXt WSL Model,”IEEE Communications Letters, vol. 25, no. 9, pp. 2953– 2957, 2021

    2021

  24. [24]

    Real-Time Radio Modulation Classification With An LSTM Auto-Encoder,

    Z. Ke and H. Vikalo, “Real-Time Radio Modulation Classification With An LSTM Auto-Encoder,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 4935–4939, 2021

    2021

  25. [25]

    Deep learning based automatic modulation recognition: Models, datasets, and challenges,

    F. Zhang, C. Luo, J. Xu, Y . Luo, and F.-C. Zheng, “Deep learning based automatic modulation recognition: Models, datasets, and challenges,” Digital Signal Processing, vol. 129, p. 103650, 2022

    2022

  26. [26]

    MR-Transformer: FPGA Acceler- ated Deep Learning Attention Model for Modulation Recognition,

    H. Wang, Z. Qi, Z. Li, and X. Zhao, “MR-Transformer: FPGA Acceler- ated Deep Learning Attention Model for Modulation Recognition,”IEEE 14 Transactions on Wireless Communications, vol. 24, no. 2, pp. 1221– 1233, 2025

    2025

  27. [27]

    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

  28. [28]

    Raw IQ Dataset for GNSS GPS Jamming Signal Classification,

    C. J. Swinney and J. C. Woods, “Raw IQ Dataset for GNSS GPS Jamming Signal Classification,”Zenodo, 2021

    2021

  29. [29]

    A Review of Recurrent Neural Net- works: LSTM Cells and Network Architectures,

    Y . Yu, X. Si, C. Hu, and J. Zhang, “A Review of Recurrent Neural Net- works: LSTM Cells and Network Architectures,”Neural Computation, vol. 31, pp. 1235–1270, 07 2019

    2019

  30. [30]

    Attention is All You Need,

    A. V . et al., “Attention is All You Need,” inAdvances in Neural Information Processing Systems(I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, eds.), vol. 30, Curran Associates, Inc., 2017

    2017

  31. [31]

    An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale,” inInternational Conference on Learning Representations, 2021

    2021

  32. [32]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Win- dows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin Transformer: Hierarchical Vision Transformer using Shifted Win- dows,” in2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9992–10002, 2021

    2021

  33. [33]

    MobileNetV2: Inverted Residuals and Linear Bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4510–4520, 2018

    2018

  34. [34]

    Searching for MobileNetV3,

    A. H. et al., “Searching for MobileNetV3,” in2019 IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), pp. 1314–1324, 2019

    2019

  35. [35]

    Automatic Modulation Recognition of Communication Signals Based on Instantaneous Statistical Character- istics and SVM Classifier,

    X. Zhang, T. Ge, and z. Chen, “Automatic Modulation Recognition of Communication Signals Based on Instantaneous Statistical Character- istics and SVM Classifier,” in2018 IEEE Asia-Pacific Conference on Antennas and Propagation (APCAP), pp. 344–346, 2018

    2018

  36. [36]

    Likelihood-Ratio Approaches to Automatic Modulation Classification,

    J. L. Xu, W. Su, and M. Zhou, “Likelihood-Ratio Approaches to Automatic Modulation Classification,”IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), vol. 41, no. 4, pp. 455–469, 2011

    2011

  37. [37]

    Online Hybrid Likelihood Based Modulation Classification Using Multiple Sensors,

    B. Dulek, “Online Hybrid Likelihood Based Modulation Classification Using Multiple Sensors,”IEEE Transactions on Wireless Communica- tions, vol. 16, no. 8, pp. 4984–5000, 2017

    2017

  38. [38]

    The Performance of Feature-based Classification of Digital Modulations Under Varying SNR and Fading Channel Conditions,

    L. Y . Uys, M. Gouws, J. J. Strydom, and A. S. J. Helberg, “The Performance of Feature-based Classification of Digital Modulations Under Varying SNR and Fading Channel Conditions,” in2017 IEEE AFRICON, pp. 198–203, 2017

    2017

  39. [39]

    Convolutional radio modulation recognition networks,

    T. J. O’Shea, J. Corgan, and T. C. Clancy, “Convolutional radio modulation recognition networks,” inProc. Engineering Applications of Neural Networks (EANN), (Cham), pp. 213–226, Springer International Publishing, 2016

    2016

  40. [40]

    Deep Neural Network Architectures for Modulation Classification,

    X. Liu, D. Yang, and A. E. Gamal, “Deep Neural Network Architectures for Modulation Classification,” in2017 51st Asilomar Conference on Signals, Systems, and Computers, pp. 915–919, 2017

    2017

  41. [41]

    An Efficient Deep Learning Model for Automatic Modulation Recognition Based on Parameter Estimation and Transformation,

    F. Zhang, C. Luo, J. Xu, and Y . Luo, “An Efficient Deep Learning Model for Automatic Modulation Recognition Based on Parameter Estimation and Transformation,”IEEE Communications Letters, vol. 25, no. 10, pp. 3287–3290, 2021

    2021

  42. [42]

    A Tim- ing Channel Spyware for the CSMA/CA Protocol,

    N. Kiyavash, F. Koushanfar, T. P. Coleman, and M. Rodrigues, “A Tim- ing Channel Spyware for the CSMA/CA Protocol,”IEEE Transactions on Information Forensics and Security, vol. 8, no. 3, pp. 477–487, 2013

    2013

  43. [43]

    Silicon Demonstration of Hardware Trojan Design and Detection in Wireless Cryptographic ICs,

    Y . Liu, Y . Jin, A. Nosratinia, and Y . Makris, “Silicon Demonstration of Hardware Trojan Design and Detection in Wireless Cryptographic ICs,” IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 25, no. 4, pp. 1506–1519, 2017

    2017

  44. [44]

    Amplitude-Modulating Analog/RF Hardware Trojans in Wireless Networks: Risks and Remedies,

    K. S. Subramani, N. Helal, A. Antonopoulos, A. Nosratinia, and Y . Makris, “Amplitude-Modulating Analog/RF Hardware Trojans in Wireless Networks: Risks and Remedies,”IEEE Transactions on In- formation Forensics and Security, vol. 15, pp. 3497–3510, 2020

    2020

  45. [45]

    Concurrent hardware Tro- jan detection in wireless cryptographic ICs,

    Y . Liu, G. V olanis, K. Huang, and Y . Makris, “Concurrent hardware Tro- jan detection in wireless cryptographic ICs,” in2015 IEEE International Test Conference (ITC), pp. 1–8, 2015

    2015

  46. [46]

    Secret Agent Radio: Covert Communication through Dirty Constellations,

    A. Dutta, D. Saha, D. Grunwald, and D. Sicker, “Secret Agent Radio: Covert Communication through Dirty Constellations,” inInformation Hiding(M. Kirchner and D. Ghosal, eds.), (Berlin, Heidelberg), pp. 160– 175, Springer Berlin Heidelberg, 2013

    2013

  47. [47]

    Practical covert channels for wifi systems,

    J. Classen, M. Schulz, and M. Hollick, “Practical covert channels for wifi systems,” in2015 IEEE Conference on Communications and Network Security (CNS), pp. 209–217, 2015

    2015

  48. [48]

    Leaking wireless ics via hardware trojan-infected synchronization,

    A. R. D ´ıaz-Rizo, H. Aboushady, and H.-G. Stratigopoulos, “Leaking wireless ics via hardware trojan-infected synchronization,”IEEE Trans- actions on Dependable and Secure Computing, vol. 20, no. 5, pp. 3845– 3859, 2023

    2023

  49. [49]

    AI-enabled covert channel detection in RF receiver architectures,

    A. E. Abdelazim, A. R. Diaz-Rizo, H. Aboushady, and H.-G. Stratigopoulos, “AI-enabled covert channel detection in RF receiver architectures,”arXiv:2604.14987, 2026

    Pith/arXiv arXiv 2026

  50. [50]

    GPS Interference Signal Recognition Based on Machine Learning,

    J. Xu, S. Ying, and H. Li, “GPS Interference Signal Recognition Based on Machine Learning,”Mob. Netw. Appl., vol. 25, p. 2336–2350, Dec. 2020

    2020

  51. [51]

    GNSS Spoofing Jamming Recognition Based on Machine Learning,

    P. Gao, S. Sun, Z. Zeng, and C. Wang, “GNSS Spoofing Jamming Recognition Based on Machine Learning,” inSignal and Information Processing, Networking and Computers(S. Sun, N. Chen, and T. Tian, eds.), (Singapore), pp. 221–228, Springer Singapore, 2018

    2018

  52. [52]

    Efficient GNSS Jamming Mitigation Using the Marcenko Pastur Law and Karhunen–Loeve Decomposition,

    O. Sharifi-Tehrani, M. F. Sabahi, and M. Raees Danaee, “Efficient GNSS Jamming Mitigation Using the Marcenko Pastur Law and Karhunen–Loeve Decomposition,”IEEE Transactions on Aerospace and Electronic Systems, vol. 58, no. 3, pp. 2291–2303, 2022

    2022

  53. [53]

    Multimodal deep neural network for UA V GPS jamming attack detection,

    F. Abdullayeva and O. Valikhanli, “Multimodal deep neural network for UA V GPS jamming attack detection,”Cyber Security and Applications, vol. 3, p. 100094, 2025

    2025

  54. [54]

    3D Radio Map- Based GPS Spoofing Detection and Mitigation for Cellular-Connected UA Vs,

    Y . Dang, A. Karakoc, S. Norshahida, and R. J ¨antti, “3D Radio Map- Based GPS Spoofing Detection and Mitigation for Cellular-Connected UA Vs,”IEEE Transactions on Machine Learning in Communications and Networking, vol. 1, pp. 313–327, 2023

    2023

  55. [55]

    A Deep Learning Based Induced GNSS Spoof Detection Framework,

    A. Iqbal, M. N. Aman, and B. Sikdar, “A Deep Learning Based Induced GNSS Spoof Detection Framework,”IEEE Transactions on Machine Learning in Communications and Networking, vol. 2, pp. 457–478, 2024

    2024

  56. [56]

    A Systematic Literature Review of Current Machine Learning Approaches for Detecting GNSS Spoofing Attacks,

    S. Bertram, L. Eisentraut, and R. Buettner, “A Systematic Literature Review of Current Machine Learning Approaches for Detecting GNSS Spoofing Attacks,”IEEE Access, vol. 13, pp. 108898–108917, 2025

    2025

  57. [57]

    EDNet: Edge- Optimized Small Target Detection in UA V Imagery - Faster Context Attention, Better Feature Fusion, and Hardware Acceleration,

    Z. Song, Y . Zhang, and A. A. R. M. A. Ebayyeh, “EDNet: Edge- Optimized Small Target Detection in UA V Imagery - Faster Context Attention, Better Feature Fusion, and Hardware Acceleration,” in21st IEEE International Conference on Ubiquitous Intelligence and Comput- ing (UIC) - 2024 IEEE Smart World Congress (SWC), pp. 829–838, 2024

    2024

  58. [58]

    A Deep Learning-Based Multi- Signal Radio Spectrum Monitoring Method for UA V Communication,

    C. Hou, D. Fu, Z. Zhou, and X. Wu, “A Deep Learning-Based Multi- Signal Radio Spectrum Monitoring Method for UA V Communication,” Drones, vol. 7, no. 8, 2023

    2023

  59. [59]

    Real-time Automatic Modulation Classification using RF- SoC,

    S. Tridgell, D. Boland, P. H. Leong, R. Kastner, A. Khodamoradi, and Siddhartha, “Real-time Automatic Modulation Classification using RF- SoC,” in2020 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pp. 82–89, 2020

    2020

  60. [60]

    FPGA-based Deep Learning Accelerator for RF Applications,

    H. den Boer, R. Muller, S. Wong, and V . V oogt, “FPGA-based Deep Learning Accelerator for RF Applications,” inMILCOM 2021 - 2021 IEEE Military Communications Conference (MILCOM), pp. 751–756, 2021

    2021

  61. [61]

    Efficient On-chip Acceleration of Machine Learning Models for Detection of RF Signal Modulation,

    J. Woo, K. Jung, and S. Mukhopadhyay, “Efficient On-chip Acceleration of Machine Learning Models for Detection of RF Signal Modulation,” in 2021 IEEE MTT-S International Microwave Symposium (IMS), pp. 74– 77, 2021

    2021

  62. [62]

    RadioML Meets FINN: Enabling Future RF Applications With FPGA Streaming Architectures,

    F. Jentzsch, Y . Umuroglu, A. Pappalardo, M. Blott, and M. Platzner, “RadioML Meets FINN: Enabling Future RF Applications With FPGA Streaming Architectures,”IEEE Micro, vol. 42, no. 6, pp. 125–133, 2022

    2022

  63. [63]

    Real-Time Automatic Modulation Recognition Based on FPGA,

    K. Zhang, B. Gao, B. Yang, Y . Ji, F. Gao, and Y . Li, “Real-Time Automatic Modulation Recognition Based on FPGA,” in2023 9th International Conference on Computer and Communications (ICCC), pp. 1440–1444, 2023

    2023

  64. [64]

    Sparsity-aware streaming snn accelerator with output-channel dataflow for automatic modulation classification,

    K. Yang, L. Zhang, A. M. Eltawil, and K. N. Salama, “Sparsity-aware streaming snn accelerator with output-channel dataflow for automatic modulation classification,”IEEE Transactions on Circuits and Systems for Artificial Intelligence, pp. 1–14, 2026

    2026

  65. [65]

    Squeeze-and-Excitation Networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-Excitation Networks,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7132–7141, 2018

    2018

  66. [66]

    Steam++ an extensible end-to-end framework for developing iot data processing applications in the fog,

    M. Gomes, R. Righi, C. Andr ´e da Costa, and D. Griebler, “Steam++ an extensible end-to-end framework for developing iot data processing applications in the fog,”International Journal of Computer Science and Information Technology, vol. 14, pp. 31–51, 02 2022

    2022

  67. [67]

    Separable Self-Attention for Mobile Vision Transformers,

    S. Mehta and M. Rastegari, “Separable Self-Attention for Mobile Vision Transformers,”Transactions on Machine Learning Research, 2023

    2023