Semantic Feature Division Multiple Access for Multi-User Digital Interference Networks

Shuai Ma, Chuanhui Zhang, Bin Shen, Youlong Wu, Hang Li, Shiyin Li, Guangming Shi, Naofal Al-DhahirIEEE Trans. Wireless Communications 2024 · DOI

TL;DR

This paper introduces Semantic Feature Division Multiple Access (SFDMA), a novel approach for multi-user communication that leverages deep learning to divide semantic information into distinct subspaces, significantly reducing interference. SFDMA enables simultaneous transmission over the same resources and is shown to improve both classification and image reconstruction tasks, outperforming existing methods, especially in low-SNR conditions.

What Problem Does This Paper Address?

Modern wireless networks face a critical challenge: the ever-increasing demand for data and the proliferation of connected devices are straining limited spectrum resources. This leads to massive access requirements and severe multi-user interference, which degrades network capacity and performance. Traditional multiple access schemes (like OMA and NOMA) have limitations in handling this complexity. While semantic communication offers a promising direction by reducing transmitted data, existing multi-user semantic communication systems are still in their infancy and struggle with interference management. Furthermore, there's a lack of theoretical understanding of how transmit power affects semantic communication performance, hindering adaptive power control strategies.

Methods

The core innovation is Semantic Feature Division Multiple Access (SFDMA). Instead of dividing resources by time, frequency, or code, SFDMA divides the semantic feature domain.

Here's how it works:

  • Semantic Feature Extraction: Deep learning models (semantic encoders) extract semantic information from user data and represent it as discrete feature vectors.
  • Feature Subspace Division: These semantic feature vectors are mapped into distinct, approximately orthogonal semantic subspaces. This separation is key to minimizing interference.
  • Simultaneous Transmission: Multiple users can then transmit their distinct semantic features simultaneously over the same time-frequency resources.
  • Interference Mitigation: At the receiver, semantic decoders are designed to effectively isolate and decode the intended semantic features, leveraging their orthogonality to suppress interference from other users.

The paper also proposes:

  • Robust Information Bottleneck (RIB) based SFDMA: For inference tasks, this approach balances informativeness, robustness, and multi-user interference by optimizing the encoded representation.
  • Swin Transformer-based SFDMA: For image reconstruction, this utilizes Swin Transformers for feature extraction and reconstruction, enabling both centralized and distributed coordinated SFDMA.
  • Alpha-Beta-Gamma (ABG) Formula: A novel theoretical relationship is established between inference accuracy and transmission power, allowing for adaptive power control strategies.

Datasets

  • MNIST: Used for classification tasks.
  • CelebFaces Attributes (CelebA): Used for image reconstruction tasks.

Key Insights and Ideas

  • Feature Domain as a Resource: The paper demonstrates that the semantic feature domain itself can be utilized as a resource for multiple access, offering a new paradigm beyond traditional resource allocation.
  • Orthogonality is Key: By mapping semantic features into approximately orthogonal subspaces, SFDMA effectively tackles multi-user interference, a major bottleneck in current systems.
  • Low-SNR Robustness: SFDMA shows a significant advantage in low-SNR regimes, outperforming even interference-free deep JSCC schemes when trained with interference. This highlights its robustness.
  • Theoretical Power-Accuracy Link: The ABG formula provides the first theoretical link between inference accuracy and transmission power in semantic communications, paving the way for intelligent power management.
  • Privacy Preservation: The distinct semantic subspaces also inherently offer a degree of privacy, as semantic information from one user is less likely to be decoded by another.

Key Results

  • Classification Accuracy: SFDMA achieves higher classification accuracy than deep JSCC with interference across various SNRs, particularly excelling at low SNRs. For example, with 64-bit quantization and 5dB SNR, SFDMA achieved accuracies of 92.79% and 92.93% for two users, significantly outperforming other methods.
  • Image Reconstruction: SFDMA demonstrates superior image reconstruction quality (measured by PSNR and MS-SSIM) compared to deep JSCC with interference, especially in low-SNR conditions.
  • Orthogonality Verification: Extensive experiments confirm the approximate orthogonality of semantic features generated by SFDMA, with inner products close to zero and angles near 90 degrees.
  • Adaptive Power Control: The proposed adaptive power control method, based on the ABG formula, significantly reduces outage probability (to 0%) compared to fixed power methods, ensuring QoS in fading channels.

Limitations and Future Work

  • Computational Complexity: While SFDMA reduces interference, the deep learning models used for semantic encoding and decoding can be computationally intensive.
  • Perfect Orthogonality: Achieving perfect orthogonality in feature subspaces is practically challenging, and the paper acknowledges that the orthogonality is approximate.
  • Generalizability: While tested on MNIST and CelebA, further validation on a wider range of datasets and tasks would be beneficial.
  • Future Work: The paper suggests exploring more advanced deep learning architectures for semantic encoders/decoders and investigating the application of SFDMA in more complex network scenarios. The established ABG formula opens avenues for further research into optimal power allocation and resource management in semantic communication systems.

Summary updated

BibTeX
@article{ma2024semantic,
  author = {Shuai Ma and Chuanhui Zhang and Bin Shen and Youlong Wu and Hang Li and Shiyin Li and Guangming Shi and Naofal Al-Dhahir},
  title = {{Semantic Feature Division Multiple Access for Multi-User Digital Interference Networks}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2024},
  doi = {10.1109/twc.2024.3427675},
}
Figure