DUAL Distributed compUting, optimizAtion, and Learning (DUAL) group at USyd

WaSeCom – Wasserstein Distributionally Robust Wireless Semantic Communication with Large AI Models

semantic-communication large-ai-models wireless-communication distributionally-robust-optimization uncertainty robustness

WaSeCom – Wasserstein Distributionally Robust Wireless Semantic Communication with Large AI Models

WaSeCom is a novel, model-agnostic framework that enhances the reliability and adaptability of wireless semantic communication (SemCom) systems against both semantic-level and channel-level uncertainties. Leveraging large AI models and Wasserstein distributionally robust optimization (WDRO), WaSeCom ensures that semantic fidelity is preserved even under adversarial perturbations, distributional shifts, and challenging wireless channel conditions.

✍️ Authors

  • Long Tan Le
  • Senura Hansaja Wanasekara
  • Zerun Niu
  • Yansong Shi
  • Nguyen H. Tran, Senior Member, IEEE
  • Phuong Vo
  • Walid Saad, Fellow, IEEE
  • Dusit Niyato, Fellow, IEEE
  • Zhu Han, Fellow, IEEE
  • Choong Seon Hong, Fellow, IEEE
  • H. Vincent Poor, Life Fellow, IEEE

🚀 Project Status: Active research with peer-reviewed publications and experimental validation across multiple modalities.

🎯 Research Focus

This project addresses core challenges in next-generation semantic communication:

  • Robust Semantic Encoding & Decoding: Extracting and reconstructing meaning with resilience to semantic noise, adversarial perturbations, and input distribution shifts.
  • Robust Channel Encoding & Decoding: Ensuring reliable transmission under unpredictable wireless channel variations (AWGN, Rayleigh fading, interference).
  • Bi-level WDRO Framework: Formulating robustness at both semantic and physical layers through Wasserstein ambiguity sets.
  • Model-Agnostic Integration: Applying the framework to large AI architectures such as BERT, Vision Transformers, and multimodal encoders.

🔬 Research Methodology

1. Semantic Communication with Large AI Models

  • Uses transformer-based encoders (e.g., BERT, ViT, wav2vec) to produce compact semantic representations.
  • Supports multiple modalities — text, images, audio, video — without modality-specific constraints.

2. Bi-level WDRO Optimization

  • Inner Level: Optimizes semantic encoder/decoder for worst-case input shifts within a semantic Wasserstein ball.
  • Outer Level: Optimizes channel encoder/decoder for worst-case channel perturbations.
  • Decouples semantic robustness from channel robustness for better generalization.

3. Dual Reformulation & Efficient Training

  • Employs Kantorovich duality to transform intractable min–max problems into scalable saddle-point optimizations.
  • Uses log-sum-exp smoothing to make worst-case optimization differentiable and compatible with deep learning.

📚 Key Contributions

  • WaSeCom Framework: First WDRO-based bi-level optimization for wireless semantic communication.
  • Formal Robustness Guarantees: Theoretical generalization bounds for both semantic and channel levels under distributional shifts.
  • Model-Agnostic Design: Applicable to various large AI model architectures and modalities.
  • Dual Robustness: Simultaneously addresses semantic and physical-layer uncertainties.
  • Empirical Validation: Outperforms state-of-the-art baselines in image and text SemCom tasks under both clean and noisy conditions.
Distributionally Robust Wireless Semantic Communication with Large AI Models
Long Tan Le, Senura Hansaja Wanasekara, Zerun Niu, Yansong Shi, Nguyen H. Tran, …, Walid Saad, Dusit Niyato, Zhu Han, Choong Seon Hong, H. Vincent Poor
arXiv  ·  01 Jan 2024  ·  doi:10.48550/ARXIV.2506.03167
A distributionally robust approach for wireless semantic communication with large AI models that addresses uncertainty in wireless channels and semantic information transmission, enhancing reliability in AI-powered communication systems.
ArXiv
semantic-communication large-ai-models distributionally-robust-optimization wireless-communication uncertainty robustness

📊 Experimental Highlights

  • Datasets: CIFAR-10 (images) & Europarl (multilingual text).
  • Channels: AWGN and Rayleigh fading, across wide SNR ranges.
  • Baselines: Compared with DeepJSCC, DeepSC, and DeepSC-RI.
  • Results:
    • Maintains high PSNR/SSIM for images under FGSM noise and channel fading.
    • Achieves consistently higher BLEU scores in text transmission, especially under adversarial and low-SNR conditions.
    • Demonstrates minimal performance loss in clean conditions while substantially improving robustness in noisy scenarios.

🏆 Research Impact

The WaSeCom framework benefits:

  • 6G and Beyond: Supports efficient, robust semantic-aware communication in next-gen wireless networks.
  • Mission-Critical Applications: Ensures reliable meaning transmission in autonomous driving, remote surgery, and industrial automation.
  • AI-Powered Systems: Enhances generalization across tasks and modalities for AI-driven communication.
  • Bandwidth Optimization: Reduces unnecessary data transmission by focusing on task-relevant semantics.

👥 Research Team

Collaboration between researchers from:

  • University of Sydney
  • Kyung Hee University
  • Virginia Tech
  • Nanyang Technological University
  • University of Houston
  • Princeton University

📧 Contact

For more information about WaSeCom, please contact the DUAL research group or the project leads.