Sub-6 GHz Over-the-Air AMC via Curriculum Fine-Tuned CNN-Transformers
Nurettin Safak, Muhammet Sefa Demirel, Alperen Marasli, Taha Eren Atmaca, Durdu Can Yerdeyatar, Ozgun Ersoy
eess.SP · cs.LG
Quick verdict
Curriculum fine-tuned CNN-Transformer keeps 91.8-93.7% OTA accuracy at 4 GHz; transparent about confounds, but no seeds, CIs, or external baselines.
Summary
The paper presents a curriculum fine-tuning study that adapts a hybrid CNN-Transformer automatic modulation classification (AMC) model from a controlled 915 MHz cable-synchronized base dataset (32 classes, ~1.85M windows) to a real 4 GHz free-space link between two USRP radios with directional antennas. Five distances (25–75 cm) were tested: three with near-ideal antenna alignment and two with deliberate partial misalignment fine-tuned last. Freezing the convolutional stem and updating only the Transformer encoder and task heads, the model retains 91.8–93.7% matched-distance test accuracy across all conditions. Ablations show the Transformer adds 2.28 points of validation accuracy over a ResNet-only variant, multi-task supervision wins at four of five distances, and stem freezing slightly reduces accuracy but cuts trainable parameters by ~44%. The authors candidly acknowledge that the sequential curriculum confounds antenna alignment with cumulative fine-tuning, that QPSK/QAM4 and DSSS/BPSK confusions stem from class taxonomy, and that results come from single runs without uncertainty estimates or external baselines.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 4
● low 0
4 strengths · 5 weaknesses · 0 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Sequential curriculum fine-tuning with a frozen 1D-ResNet stem keeps matched-distance test accuracy within 91.8–93.7% across five 4 GHz free-space conditions, including two partially misaligned ones fine-tuned last.
- The Transformer encoder contributes 2.28 points of Stage 0 validation accuracy over a ResNet-only variant (94.60% vs 92.32%).
- Multi-task supervision outperforms single-task supervision at four of five distances (92.74% vs 92.19% average).
- Freezing the stem slightly reduces accuracy (92.74% vs 93.09% no-freeze) but cuts updated parameters per fine-tuning step by ~44%, retained as a compute trade-off.
- The dominant QPSK/QAM4 and DSSS/BPSK confusions are class-taxonomy and channel-physics artifacts present at comparable rates in all five conditions, not classifier failures.
Strengths
Real over-the-air evaluation, not simulation
Section III and Table I evaluate on a genuine 4 GHz free-space link (USRP N-series pair, log-periodic UWB antennas) rather than synthetic or cabled data, directly probing the sim-to-real gap the introduction motivates via O'Shea et al.
Unusual transparency about its own confound
Section IV/V-A explicitly state that curriculum order confounds alignment with cumulative fine-tuning and decline to interpret Stage 2's higher numbers (93.67%/93.11%) as misalignment robustness; Table IV reports the failed freeze-stem hypothesis rather than hiding it
Physically grounded confusion analysis
Three design-choice ablations
Weaknesses
Curriculum design confounds the key variables
Alignment manipulation never quantified
Single-run results without uncertainty
Limited link severity and unknown SNRs
No external baselines
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper (arXiv:2609.07726) targets the sim-to-real gap in automatic modulation classification (AMC). A hybrid CNN-Transformer (Plan A: 1D-ResNet stem reducing 1024 samples to 64 timesteps at d_model=128; 3-layer/4-head Transformer encoder; five task heads with homoscedastic uncertainty weighting; 793,265 parameters) is base-trained at 915 MHz on a clock/PPS-synchronized cabled SISO link (32 classes, ~1.85M windows), then sequentially fine-tuned to a real 4 GHz free-space link at five distances (25/35/50/70/75 cm), three ≈99% antenna-aligned and two ≈85% aligned. Central claims: (1) matched-distance accuracy stays in 91.8–93.7% across all five conditions (Table I); (2) the Transformer encoder adds 2.28 validation points over a ResNet-only variant (Table II); (3) multi-task supervision wins at 4/5 distances (Table III); (4) stem freezing trades <0.4 accuracy points for ~44% fewer updated parameters (Table IV); (5) QPSK/QAM4 and DSSS/BPSK confusions are taxonomy artifacts. The authors repeatedly flag their own confound: curriculum order makes distance, alignment, and cumulative adaptation non-independent.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 0 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy does the model achieve over the air?
Matched-distance test accuracy was 91.8–93.7% across the five 4 GHz conditions, each evaluated on 192,000 test windows: 92.23% at 25 cm, 92.86% at 50 cm, and 91.81% at 75 cm with near-ideal antenna alignment (Stage 1), and 93.67% at 35 cm and 93.11% at 70 cm with deliberate partial misalignment (Stage 2). The authors caution that the sequential fine-tuning order confounds alignment with cumulative adaptation, so the higher Stage 2 numbers should not be read as evidence that misalignment helps.
Why is QPSK confused with QAM4 in this model?
QPSK and 4-QAM (4-QAM) describe the same four-point constellation under different naming conventions, so from raw I/Q samples alone they are physically indistinguishable. In this paper QPSK was predicted as QAM4 in 4282 of 6000 test windows at 25 cm, and the authors argue the confusion reflects the class taxonomy rather than a classifier deficiency; DSSS shows the analogous issue against BPSK because the DSSS signals are BPSK-chip modulated.
What does freezing the CNN stem do during fine-tuning?
At each fine-tuning step only the Transformer encoder and the five task heads were updated while the 1D-ResNet stem stayed frozen, cutting updated parameters by roughly 44% per step. Ablation showed this slightly lowered average matched-distance accuracy (92.74% vs 93.09% without freezing), so the authors retain it as a compute trade-off for resource-constrained onboard deployment, not as an accuracy improvement.
Was the MIMO expansion cable used to transmit the data?
No. The cable only shared a reference clock and PPS timing signal between the two USRP radios for coherent synchronization; the RF link itself was single-input single-output (SISO). All 4 GHz experimentation — fine-tuning and every reported over-the-air accuracy figure — used genuine free-space transmission through directional log-periodic antennas.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (confusion matrices, I/Q constellation panels, setup photos) are images absent from the extracted text; only figure-internal numbers quoted in prose (e.g., 4282/6000 QPSK→QAM4 at 25 cm) could be verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.