GenSC-6G: A Prototype Testbed for Integrated Generative AI, Quantum, and Semantic Communication

TL;DR

GenSC-6G addresses the need for integrated generative AI, quantum processing, and semantic communication in 6G by providing a modular testbed and a noise-augmented semantic dataset. The approach combines diffusion-driven data generation, bandwidth-efficient semantic compression, and JSCC over both classical and quantum channels, supported by HQC parallel processing. Key contributions include an adaptable SC framework, a diffusion-based auto dataset pipeline with automatic labeling, mobile-DL support, and case studies on lightweight classification, semantic localization, upsampling recovery, and edge LLMs. The results demonstrate robustness and scalability under varied $SNR$ conditions, highlighting the framework's potential to enable energy-efficient, context-aware semantic communications for future networks.

Abstract

We introduce a prototyping testbed, GenSC-6G, developed to generate a comprehensive dataset that supports the integration of generative artificial intelligence (AI), quantum computing, and semantic communication for emerging sixth-generation (6G) applications. The GenSC-6G dataset is designed with noise-augmented synthetic data optimized for semantic decoding, classification, and localization tasks, significantly enhancing flexibility for diverse AI-driven communication applications. This adaptable prototype supports seamless modifications across baseline models, communication modules, and goal-oriented decoders. Case studies demonstrate its application in lightweight classification, semantic upsampling, and edge-based language inference under noise conditions. The GenSC-6G dataset serves as a scalable and robust resource for developing goal-oriented communication systems tailored to the growing demands of 6G networks.

Figures (4)

• Figure 1: GenSC-6G dataset structuress. The dataset consists of the ground-truth data, encoded features, and additive noise. The AI-6G use cases span narrow AI and generative AI applications over SC and goal-oriented tasks in multiple fields. ResNet, ViT, SwinT, AWGN, RFI, LLM, and RIS stand for residual network, vision transformer, swin transformer, additive white Gaussian noise, radio frequency interference, large language model, and reconfigurable intelligent surface, respectively.
• Figure 2: A GenSC-6G testbed prototype. The large-AI SC testbed framework prototypes a flexible architecture in which the backbone encoder and communication modules are alterable to fit any backend, and the semantic decoders can be adapted for various downstream goal-oriented tasks. ReLU, TX, RX, mmWave, OFDM, PSK, QAM, MHSA, BLIP, GPT, LLaMA, and FPN stand for the rectified linear unit, transmitter, receiver, millimeter wave, orthogonal frequency division multiplexing, phase-shift keying, quadrature amplitude modulation, multi-head self-attention, bootstrapping language-image pretraining, generative pretrained transformer, LLM Meta AI, and feature pyramid network, respectively.
• Figure 3: A transceiver setup to capture noise features as part of the testbed. The testbed leverages the Wi-Fi $7$ ($802.11$be) OFDM communication system with file streaming. On the transmitter side, a programmable SDR setup with gigahertz (GHz) antennas and amplifiers sends high-frequency signals to the receiver.
• Figure 4: Overview of case studies demonstrating downstream goal-oriented tasks from encoded features and their common metrics, including semantic compression, object localization, recovery through upsampling and diffusion, and post-processing with LLM, illustrating the adaptability of encoded features. Herein, the average compression rate of the dataset reduces to $99.993\%$, which is compressed to nearly zero. The confusion matrix (bottom left) illustrates the classification performance of the ResNet-50 model under different SNR conditions. The LPIPS probability distribution and the PSNR (bottom right) contrast perceptual similarity scores between different models and reflect image quality maintenance across varying noise levels, showing the adaptability and effectiveness of these models.