Mod-CL: Modulation Consistency-based Contrastive Learning for Self-Supervised Automatic Modulation Classification

This repository provides the official implementation of the paper:

Modulation Consistency-based Contrastive Learning for Self-Supervised Automatic Modulation Classification

📄 Paper: arXiv:2605.11875

Overview

Mod-CL is a self-supervised contrastive learning framework for automatic modulation classification (AMC).

The key idea is to exploit intra-instance modulation consistency: different temporal segments of the same signal instance may have different local waveform patterns, but they share the same modulation type.

Based on this observation, Mod-CL constructs positive pairs from different temporal segments of the same IQ signal and learns modulation-discriminative representations without requiring labels during pretraining.

Framework

During self-supervised pretraining, each IQ signal is:

1. Augmented into two stochastic views;
2. Split into temporal segments;
3. Encoded by a shared encoder;
4. Optimized using a modulation-consistent contrastive loss.

The total loss consists of three parts:

• Segment Consistency Loss
• Augmentation Consistency Loss
• Joint Consistency Loss

After pretraining, the encoder is frozen and evaluated using linear probing for downstream modulation classification.

Datasets

Experiments are conducted on public RadioML datasets:

• RadioML 2016.10A
• RadioML 2016.10B

Please place the datasets under the data/ directory.

data/
├── RML2016.10a/
└── RML2016.10b/

Code

Code is currently under preparation and will be released soon.

Main Results

We evaluate Mod-CL on two public automatic modulation classification benchmarks, RadioML 2016.10A and RadioML 2016.10B.

Following the standard linear probing protocol, the encoder is first pretrained without labels and then frozen. Only a linear classifier is trained using limited labeled samples. Here, N denotes the number of labeled samples per modulation class per SNR level.

Linear Probing Results

Mod-CL consistently outperforms generic self-supervised methods and AMC-oriented self-supervised baselines under all label budgets.

RadioML 2016.10A

Method	N=2	N=5	N=10	N=20	N=50	N=100
Random Init.	13.59	16.08	16.95	18.23	20.13	21.52
SimCLR	45.52	47.95	49.07	50.37	51.68	52.86
MoCo	32.91	38.60	41.30	42.83	44.32	45.56
SemiAMC	36.32	43.82	47.51	49.91	53.90	56.03
MAC	27.92	35.40	41.65	47.76	54.53	58.13
Mod-CL	51.76	57.33	59.10	60.32	61.40	61.88

RadioML 2016.10B

Method	N=2	N=5	N=10	N=20	N=50	N=100
Random Init.	13.37	15.88	17.18	18.26	20.05	21.51
SimCLR	47.03	49.16	50.66	51.89	53.23	54.34
MoCo	38.07	42.25	43.44	44.13	45.44	46.55
SemiAMC	31.61	39.99	44.14	48.24	52.62	55.52
MAC	20.63	26.92	32.73	38.82	51.00	57.51
Mod-CL	53.22	58.09	60.36	61.80	62.99	63.65

These results show that Mod-CL is particularly effective in low-label regimes. For example, on RadioML 2016.10A, Mod-CL achieves 57.33% accuracy with only N=5 labeled samples, outperforming SimCLR by 9.38 percentage points. On RadioML 2016.10B, Mod-CL achieves 58.09% accuracy under the same label budget.

Per-SNR Performance

We further evaluate the performance under different SNR levels with N=5 labeled samples per modulation class per SNR.

The per-SNR results show that Mod-CL achieves the best overall trend on both datasets and maintains clear advantages in most SNR regions, especially under low-to-medium SNR conditions.

Figure: Per-SNR test accuracy under the linear probing protocol with N=5 labeled samples per class per SNR on RadioML 2016.10A and RadioML 2016.10B.

Summary

The main observations are:

• Mod-CL achieves the best linear probing accuracy across all label budgets on both RadioML 2016.10A and RadioML 2016.10B.
• The advantage is most significant when labeled data are scarce.
• Mod-CL performs consistently well across different SNR levels.

Citation

If you find this repository useful, please cite our paper:

@article{wang2026modcl,
  title={Modulation Consistency-based Contrastive Learning for Self-Supervised Automatic Modulation Classification},
  author={Wang, Chenxu and Wang, Shuang and Han, Lirong and Hu, Xinyu and Mo, Hanlin and Xing, Hantong and Tegos, Sotiris A. and Jiao, Licheng},
  journal={arXiv preprint arXiv:2605.11875},
  year={2026}
}