acceptodds
Under review as a conference paper at ICLR 2027

Beyond White Noise: Integrating Spectral Gaussian Process into Diffusion Models for Wireless Signal Recognition Pre-training

Abstract

Wireless signal recognition (WSR) aims to infer attributes of passively received wireless signals, and is widely used in radio system. Existing WSR models can learn dependencies among discrete wireless signal samples implicitly, but generally do not make the signal correlation structure an explicit target or prior during representation learning. Consequently, these dependencies must be discovered from the training data and architectural bias alone, without a direct covariance-aware learning objective. Therefore, we propose Spectral Gaussian Process Diffusion (SGP-Diff), a self-supervised framework that replaces isotropic white noise with length-scale-conditioned spectral GP noise that couples neighboring frequency bins. The model predicts this structured noise with a covariance-normalized denoising objective. A linear second-order analysis shows how the optimal denoiser depends jointly on the data and noise covariances. Across four datasets covering three WSR tasks, SGP-Diff obtains the best reported accuracy among the compared methods and exceeds the strongest baseline by 1.83% on average. Further analysis results have validated the effectiveness and practical efficiency of the designed module.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Diffusion, Wireless Signal Recognition

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_8ptfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary
  • The paper proposes SGP-Diff, a self-supervised pretraining framework for wireless signal recognition that replaces isotropic Gaussian noise with spectral Gaussian process perturbations to introduce correlations across neighboring frequency bins.
  • The method combines length-scale-conditioned denoising with a covariance-normalized loss. A linear analysis characterizes how the optimal denoiser depends on both signal and perturbation covariances.
  • Experiments cover four datasets spanning modulation classification, wireless technology classification, and transmitter identification. Results show improvements over the evaluated baselines, supported by limited-label experiments, component ablations, and efficiency analysis, although the gains attributable specifically to spectral GP perturbations are modest in the reported ablations.
Critical strengths
  • A principled formulation of structured-noise pretraining. The method explicitly incorporates spectral perturbation covariance through length-scale conditioning and a covariance-normalized objective. The linear analysis clarifies how the denoiser depends on both signal and noise covariances, providing a mathematical rationale for the design.

  • Promising performance under limited supervision. SGP-Diff outperforms TS-DDAE across all reported labeling ratios on three datasets, with particularly substantial gains on HisarMod2019. These results support the framework’s potential value when labeled wireless signals are scarce, although they do not isolate the contribution of spectral GP noise.

Critical weaknesses
  • The empirical evidence does not sufficiently establish the contribution of spectral GP perturbations. Although the full framework outperforms the evaluated baselines, replacing spectral GP with isotropic Gaussian noise under the same architecture reduces accuracy by only 0.30–0.42 percentage points across the three ablated datasets. Without results across random seeds, the reliability of these small gains remains unclear. The corresponding ablation is also missing for LoRa-RFFI. Consequently, the results do not convincingly establish a substantial advance attributable to the central proposed component or explain which signal characteristics make its covariance prior beneficial.

  • Missing comparison with a closely related method. The paper discusses PWC-Diff, which incorporates physical wireless channel priors into diffusion-based representation learning, but does not include it in the experimental comparisons or explain its omission. A direct comparison would help establish whether spectral GP perturbations offer complementary benefits or advantages over existing approaches that incorporate wireless-domain priors.

  • The practical problem setting and significance require stronger justification. For wireless technology identification, recognizable preambles and protocol structures may already support identification under passive reception. The paper does not sufficiently clarify when such information is unavailable or unreliable, or compare against relevant protocol-aware alternatives. While this concern does not directly resolve modulation classification or transmitter fingerprinting, the lack of clearly specified deployment assumptions makes it difficult to assess the framework’s practical advantage.

Questions for authors
  • Empirical reliability: In Table 3, spectral GP perturbations improve accuracy over isotropic Gaussian noise by only 0.30–0.42 percentage points. Are these gains consistent across random seeds? Please provide variability estimates and, if available, the corresponding Gaussian-noise ablation on LoRa-RFFI.

  • Comparison with prior work: Why is PWC-Diff omitted from the experimental comparisons despite being discussed as closely related work? Is there evidence that SGP-Diff offers advantages over this method under a comparable evaluation protocol?

  • Justification of the spectral prior: What evidence supports applying a Matérn covariance prior to the real and imaginary spectral coefficients, rather than only to the spectral envelope? Can the authors clarify which signal characteristics make this prior beneficial and whether those characteristics explain the dataset-dependent gains?

  • Practical problem setting: For wireless technology classification, does the evaluation assume that recognizable preambles or protocol structures are unavailable or unreliable? Please clarify the intended reception conditions and why protocol-aware identification would be insufficient in that setting.

Reviewer_PUV4OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the problem of wireless signal recognition pretraining. The paper presents SGP-Diff, a self-supervised pretraining framework for wireless signal recognition. The framework replaces independent Gaussian corruption with Matern Gaussian-process noise applied across frequency bins. Besides, it uses a dual-domain encoder to predict the injected noise while conditioning on diffusion time and kernel length scale. Experiments across four datasets spanning modulation classification, technology classification, and transmitter identification demonstrate slightly higher accuracy.

Critical strengths
  • The paper provides a self-supervised framework, SGP-Diff.
  • It develops a length-scale-conditioned denoising scheme with a covariance-normalized objective.
  • The authors evaluate SCG-Diff on four datasets to show the accuracy, data efficiency, design choices, and computational costs.
Critical weaknesses
  • The paper reports that SCG-Diff achieves a higher accuracy than the best existing approaches. However, the improvement is 1.83%, which is very marginal. As a consequence, the necessity of developing SCG-Diff is questionable.
  • It is not clear why it replaces isotropic white noise with length-scale-conditioned spectral GP noise. Wireless signals naturally are corrupted by noises. However, whether the noise follows Gaussian distributions or other distributions needs experiments to assess. We cannot directly argue that the noise follows some types of distributions.
  • The paper does not clear present the proposed SCG-Diff. Why is a diffusion model used to perform wireless signal recognition?
  • Usually, white noise is ideally unpredictable as each sample is independent of previous samples. The paper considers white noise instead of GP noise (should be structured noise). In my opinion, it forges a research question but not the question that comes from practical wireless systems.
Questions for authors
  1. Please provide the justification for the spectral covariance.
  2. Why does the paper employ a diffusion model to perform denoising?
  3. How is wireless signal recognition performed?
  4. Please clarify the measurements showing that the raw real and imaginary Fourier coefficients exhibit covariance compatible with the proposed kernel.
Reviewer_1rNMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a self-supervised pre-training framework SGP-Diff for WSR. It replaces isotropic white noise with length-scale-conditioned spectral GP noise. The model predicts this structured noise with a covariance-normalized denoising objective. A linear analysis describes the roles of signal and perturbation covariances. Experiments cover four datasets spanning modulation classification, wireless technology classification, and radio-frequency fingerprint identification.

Critical strengths
  1. This paper considers perturbation covariance as an explicit design variable. The kernel-based formulation provides a controlled method to alter the noise correlation, with a denoising loss matched to the prescribed covariance.

  2. Experiments cover four datasets and three wireless recognition tasks, with additional limited label comparisons and ablations. This provides a useful evaluation that goes beyond a single benchmark.

Critical weaknesses
  1. This paper designs relevant noise based on the local smoothness of wireless spectrum, but the authors' explanation of why "adding this special correlated noise" can help the model learn more useful recognition features is not clear enough. Is this design truly utilizing existing wireless spectrum-related structures, or is it artificially constructing a training task that is more conducive to representation learning?

  2. Further explanation is needed on the specific wireless features utilized in this design, as well as how to distinguish the effects of these features from the gains in general situations.

  3. Compared with a fixed length scale (3.1) or randomly sampled length scales without explicit conditioning (as input, 3.2), what benefits does this additional design provide for downstream recognition? This comparison would help distinguish the effects of perturbation diversity (3.1) from those of the conditioning input (3.2).

Questions for authors
  1. In Table 3, the proposed method only improved by about 0.4% compared to the white noise (w/ Gaussian noise), but there is a lack of explanation for the fluctuations in repeated experiments, and it is uncertain whether this gain is stable.

  2. This ablation comparison is also lacking in meaningful limited-label experiments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.