GenSC-6G - Scalable Semantic Communication Framework and Dataset

This repository contains the first semantic communication dataset and playground, designed to be scalable, reproducible, and adaptable for a wide range of applications. The dataset and framework are tailored for semantic decoding, classification, and localization tasks in 6G applications, integrating generative AI and semantic communication. Implementation of GenSC-6G: A Prototype Testbed for Integrated Generative AI, Quantum, and Semantic Communication.

---

Features of the GenSC-6G Dataset

🔧 Adaptable SC Framework

A flexible prototype that supports modifications to baseline models, communication modules, and decoders, enabling customization for diverse communication needs.

🤖 Generative AI-Driven SC

The integration of generative AI for synthetic data generation, enriching the Knowledge Base (KB) and leveraging large language model (LLM) capabilities for enhanced semantic tasks.

📊 Noise-Augmented Dataset

A labeled dataset with injected noise, specifically optimized for semantic tasks such as target recognition, localization, and recovery. The dataset comprises 4,829 training and 1,320 testing instances across 15 classes of military and civilian vehicle types. It incorporates Additive White Gaussian Noise (AWGN) and Radio Frequency (RF) interference at varying Signal-to-Noise Ratios (SNRs) to evaluate model robustness under realistic channel conditions.

📥 Dataset Download and Overview

Main Dataset

Download the main dataset here

Segmentation Dataset

Download the segmentation dataset here

📝 Case Study on Semantic Tasks

A detailed case study that evaluates baseline models across various semantic tasks, assessing performance and adaptability under different noise conditions to validate the GenSC-6G framework.

---

Table of Supported Case Study

Classification Models	Segmentation Models	Upsampling Models	EdgeLLM Models
ClassicalViT-L-32	UNet	ResNet-50	Llama-3
ViT-L-32	EfficientNet	DINO-V2	BLIP-2
ResNet-50	SAM	ViT-L-32	GPT-4
VGG-16			Qwen2-VL
Inception-V3			Phi3-Vision
EfficientNet-B1
MobileNet-V3

---

Setup Instructions

1. Environment Setup

• Install Anaconda.
• Create an environment using:

  conda env create -f environment.yml  # For classical-only setup
  conda env create -f environment-quantum.yml  # For HQC setup
  conda activate gensc

2. Dataset Setup

• Download the dataset from HuggingFace🤗:

  cd GenSC-Testbed
  git clone https://huggingface.co/datasets/CQILAB/GenSC-6G
  cd ..

3. Training Scripts

• For classification tasks:

  python train.py  # For quantum-based training
  python train-nonquantum.py  # For classical training

• For upsampling tasks:

  python train_super_res.py

Example of Running Experiments

• Train-test different models classification ([inceptionv3, resnet, qresnet, vit, swin, mobilenet, efficientnet, vgg, qcnn]) using:

  python train.py --model_name resnet --batch_size 32 --snr 10

  python train_nonquantum.py --model_name resnet --batch_size 32 --snr 10

• Train-test different models upsampling with:

  python train_super_res.py --model_name resnet --batch_size 32 --snr 10

For Interactive Examples

• playground-training-classification.ipynb
• playground-training-upsampling-and-edgellm.ipynb

Set the kernel to gensc.

---

Code Customization

• Encoders and Decoders:

  • Modify allmodels.py to use custom encoder networks.
  • Pair with decoders defined in decoder.py.

• Edge LLM Models:

  • Replace the edge-based large language model (LLM) with alternatives in playground-training-upsampling-and-edgellm.ipynb.

---

Reproducibility

🗃️ Dataset

Labeled dataset with ground-truth data, noise features, and extracted semantic features. Uploaded to HuggingFace🤗

Dataset Columns and Descriptions

• image: Raw image data used for training and evaluation.
• image_path: Path to the corresponding image file.
• classification_class: Integer label corresponding to the classification category (0-15).
• classification_{basemodel}_features: Extracted feature embeddings from {basemodel}'s encoder, consisting of 1000 float32 tensors.
• classification_awgn10dB_{basemodel}_features: Feature embeddings extracted from {basemodel} encoder with Additive White Gaussian Noise (AWGN) at 10dB SNR.
• classification_awgn30dB_{basemodel}_features: Feature embeddings extracted from {basemodel} encoder with AWGN at 30dB SNR.
• upsampling_{basemodel}_features: Extracted feature embeddings for upsampling tasks using {basemodel} encoder, consisting of 1000 float32 tensors.
• upsampling_awgn10dB_{basemodel}_features: Upsampling features with AWGN at 10dB SNR for {basemodel}.
• upsampling_awgn30dB_{basemodel}_features: Upsampling features with AWGN at 30dB SNR for {basemodel}.

🏗️ Testbed

To experiment with real-world semantic communication, you can use the GNURadio and HackRF.

1. Install Dependencies:
   • Install GNU Radio
   • Install HackRF tools: sudo apt install hackrf
2. Configure Transceiver:
   • Transmitter config: GNURadio/transmitter.grc
   • Outputs a streaming binary file
3. Run Transmitter:
   • Open GNURadio/transmitter.grc in GNU Radio Companion
   • Set SDR parameters (frequency, gain, bandwidth)
   • Execute to start transmission
4. Run Receiver:
   • Modify GNURadio/receiver.grc settings
   • Run to capture and process signals By following these steps, you can replicate real-world transmission experiments using the testbed and analyze its performance.

💻 Flexible Code

Modular structure for customization, including:

• Base models
• Communication modules
• Decoders
• Base model and decoder that is task awareness and potentially support more AI downstream task.

📊 Performance Metrics

• The model outputs logs in logs/ directory.

• Metrics include:

  • Accuracy
  • F1 Score
  • PSNR (Peak Signal-to-Noise Ratio)
  • SSIM (Structural Similarity Index Measure)
  • LPIPS (Learned Perceptual Image Patch Similarity)
  • CLIP-S (Contrastive Language-Image Pretraining Score for LLMs)
  • BERT Score
  • BLEU Score
  • Word Error Rate (WER)

  These metrics provide insight into the robustness of models under different noise conditions and evaluate text-image alignment for semantic tasks. For additional evaluation metrics, refer to TorchMetrics.

🔄 Model Checkpoints

• Trained models are saved in the logs/ directory. (Can be opened with tensorboard)
• Checkpoints can be loaded for continued training or evaluation.

---

Citation

The paper can be found at arXiv.

If you use this dataset or framework in your research, please cite:

@article{gensc6g,
      title={GenSC-6G: A Prototype Testbed for Integrated Generative AI, Quantum, and Semantic Communication}, 
      author={Brian E. Arfeto and Shehbaz Tariq and Uman Khalid and Trung Q. Duong and Hyundong Shin},
      year={2025},
      eprint={2501.09918},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2501.09918}, 
}

---

Others

• Official Repository: CQILAB/GenSC-6G
• Dataset: HuggingFace

License

MIT License

Contributor

• Brian Estadimas
• Shehbaz Tariq

You can contribute by adding your own model or modifying the code. Make a pull request (PR), and once it is merged, your name will be listed here.

Changelogs

• Dataset uploaded to HuggingFace and Repo intiated