Benchmarks for Automatic Modulation Recognition using Deep learning methods

1.  Deep Learning Based Automatic Modulation Recognition: Models, Datasets, and Challenges: https://github.com/Richardzhangxx/AMR-Benchmark

This repository provides the source code for the paper “Deep Learning Based Automatic Modulation Recognition: Models, Datasets, and Challenges,” published in Digital Signal Processing. It implements representative and state-of-the-art deep learning models for AMR across four widely used datasets: RML2016.10a, RML2016.10b, RML2018.01a, and HisarMod2019.1. The project serves as a unified benchmark and reference framework for researchers in the AMR field.

2.   AI Framework of Radio Recognition : https://github.com/Singingkettle/ChangShuoRadioRecognition

CSRR (ChangShuoRadioRecognition) is an open-source Automatic Modulation Classification (AMC) framework built on PyTorch and MMEngine. It supports over 20 state-of-the-art AMC algorithms and uses a config-driven design for flexible experiment management. The framework also provides built-in tools for performance analysis, including generation of publication-ready figures and tables, while maintaining minimal dependencies.

3. ChangShuo Radio Data (CSRD): https://github.com/Singingkettle/ChangShuoRadioData

ChangShuo Radio Data (CSRD) is a MATLAB-based simulation framework designed for wireless communication system analysis. It provides a comprehensive environment for generating and studying radio communication data, supporting research and development in modulation recognition and related tasks.

4. IQFormer: A Novel Transformer-Based Model with Multi-modality Fusion for Automatic Modulation Recognition: https://github.com/WestdoorSad/IQFormer

IQFormer is a Transformer-based model designed for Automatic Modulation Recognition (AMR) that incorporates multi-modality fusion to improve performance. The repository provides the official implementation of the proposed method, enabling researchers to explore how combining different signal representations enhances classification accuracy in wireless communication tasks.

5. STF-GCN: A Multi-Domain Graph Convolution Network Method for Automatic Modulation Recognition via Adaptive Correlation: https://github.com/WestdoorSad/STF-GCN

STF-GCN: A Multi-Domain Graph Convolution Network Method for Automatic Modulation Recognition via Adaptive Correlation. Official Code for "STF-GCN: A Multi-Domain Graph Convolution Network Method for Automatic Modulation Recognition via Adaptive Correlation".

6. RML22 Dataset, Dataset generation code for modulation classification: https://github.com/venkateshsathya/RML22

The RML22 repository provides code for generating a modulation classification dataset along with a complete pipeline for training and evaluation. It includes scripts to create the dataset using signal sources (e.g., audio input for analog modulation), train a CNN-based model, and evaluate performance through metrics such as accuracy versus SNR and confusion matrices. The project serves as a practical framework for dataset generation and baseline model experimentation in AMR.

7.  Convolutional Neural Network Assisted Transformer for Automatic Modulation Recognition under Large CFOs and SROs: https://github.com/DTMB-DL/TransGroupNet

TransGroupNet is a CNN-assisted Transformer model designed for Automatic Modulation Recognition (AMR) under challenging conditions such as large carrier frequency offsets (CFOs) and sampling rate offsets (SROs). The repository provides implementation code, including instructions for dataset preparation, model training, and evaluation. It enables users to train models from scratch, test performance, and analyze results through generated outputs and visualizations.

8. Adversarial Robust ViT-based Automatic Modulation Recognition in Practical Deep Learning-based Wireless Systems: https://github.com/coulsonlee/Robust-ViT-for-AMR-SP2025

A robust Automatic Modulation Recognition (AMR) system based on a Vision Transformer (ViT) is proposed to defend against adversarial attacks in non-cooperative wireless environments. The model integrates a feature extraction module with tailored feature and positional embeddings to enhance robustness.The framework consists of two main stages. First, the ViT model improves classification accuracy by capturing long-range dependencies in signals. Second, adversarial training is applied using noise-adaptive perturbations that consider real-world noise conditions, making the generated adversarial samples more practical and effective.Experimental results on a self-collected real-world dataset demonstrate the efficiency, accuracy, and robustness of the proposed approach.

9.