Distributed Deep Joint Source-Channel Coding over a Multiple Access Channel

Selim F. Yilmaz, Can Karamanli, Deniz Gündüz

Abstract

We consider distributed image transmission over a noisy multiple access channel (MAC) using deep joint source-channel coding (DeepJSCC). It is known that Shannon's separation theorem holds when transmitting independent sources over a MAC in the asymptotic infinite block length regime. However, we are interested in the practical finite block length regime, in which case separate source and channel coding is known to be suboptimal. We introduce a novel joint image compression and transmission scheme, where the devices send their compressed image representations in a non-orthogonal manner. While non-orthogonal multiple access (NOMA) is known to achieve the capacity region, to the best of our knowledge, non-orthogonal joint source channel coding (JSCC) scheme for practical systems has not been studied before. Through extensive experiments, we show significant improvements in terms of the quality of the reconstructed images compared to orthogonal transmission employing current DeepJSCC approaches particularly for low bandwidth ratios. We publicly share source code to facilitate further research and reproducibility.

Introduction

Traditional wireless communication systems are built on the principle of separation: source coding compresses data, while channel coding adds redundancy for error protection. While Shannon's theory proves this separation is optimal for infinite block lengths, it often fails in practical scenarios involving finite-length data, such as individual images, especially under low signal-to-noise ratio (SNR) conditions.

The challenge intensifies in a Multiple Access Channel (MAC) environment where multiple devices share a single receiver. Conventional approaches rely on orthogonal transmission methods like Time Division Multiple Access (TDMA), where users take turns. However, information theory suggests that non-orthogonal methods—where signals are superposed—can achieve higher efficiency. The primary hurdle is designing a practical system capable of "untangling" these superposed complex signals. This paper introduces DeepJSCC-NOMA, a framework that combines Non-Orthogonal Multiple Access (NOMA) with joint source-channel coding to address these limitations.

Proposed Architecture: DeepJSCC-NOMA

The proposed DeepJSCC-NOMA architecture leverages deep neural networks (DNNs) to map image pixels directly to complex-valued channel symbols, bypassing the traditional separation of source and channel coding.

The DeepJSCC-NOMA system model featuring Siamese encoders and a shared wireless MAC.

Key technical innovations include:

  • Siamese Network Architecture: To ensure scalability and reduce parameter overhead, all transmitters utilize identical encoder parameters. This prevents parameter explosion as the number of users increases.
  • Device Embeddings: To resolve the identity of superposed signals at the receiver, unique trainable 1D "device embeddings" are concatenated with the input images, acting as digital fingerprints that help the decoder separate the signals.
  • Curriculum Learning (CL): Training the decoder to separate interfering signals is complex. A progressive training strategy is employed, starting with interference-free transmission and gradually introducing the full NOMA task.
  • Power Normalization: A dedicated layer ensures each transmitter adheres to its specific power budget.

Experimental Results and Discussion

The framework was evaluated using the CIFAR-10 dataset (60,000 32x32 RGB images). The performance of DeepJSCC-NOMA was compared against a TDMA-based DeepJSCC baseline across various SNR levels and bandwidth ratios ($\rho$).

PSNR performance comparison of DeepJSCC-NOMA-CL against TDMA and single-user baselines.

Key Findings

  • Superior Reconstruction Quality: For a bandwidth ratio of $\rho = 1/3$, the NOMA method with Curriculum Learning (NOMA-CL) achieved a 0.91 dB improvement in Peak Signal-to-Noise Ratio (PSNR) over the TDMA baseline. At a tighter ratio of $\rho = 1/6$, the improvement remained significant at 0.25 dB.
  • Graceful Degradation: Unlike traditional digital systems that suffer from the "cliff effect" (sudden failure as noise increases), DeepJSCC-NOMA exhibits graceful degradation, maintaining decodable image quality even as channel conditions worsen.
  • Fairness and Efficiency: The system achieves nearly identical PSNR for both users, demonstrating fairness. Furthermore, the Siamese architecture allows for multi-user scaling with only a marginal increase in computational and memory complexity compared to a single-user system.

Conclusion and Future Work

DeepJSCC-NOMA demonstrates that non-orthogonal superposition, combined with deep learning-based joint source-channel coding, significantly outperforms traditional orthogonal methods in the finite block length regime. While current results on CIFAR-10 are promising, future research will focus on scaling the system to higher-resolution datasets like ImageNet, increasing the number of simultaneous users, and adapting the model to more complex real-world channel conditions involving fading and synchronization offsets.

Summary updated

Abstract

We consider distributed image transmission over a noisy multiple access channel (MAC) using deep joint source-channel coding (DeepJSCC). It is known that Shannon's separation theorem holds when transmitting independent sources over a MAC in the asymptotic infinite block length regime. However, we are interested in the practical finite block length regime, in which case separate source and channel coding is known to be suboptimal. We introduce a novel joint image compression and transmission scheme, where the devices send their compressed image representations in a non-orthogonal manner. While non-orthogonal multiple access (NOMA) is known to achieve the capacity region, to the best of our knowledge, non-orthogonal joint source channel coding (JSCC) scheme for practical systems has not been studied before. Through extensive experiments, we show significant improvements in terms of the quality of the reconstructed images compared to orthogonal transmission employing current DeepJSCC approaches particularly for low bandwidth ratios. We publicly share source code to facilitate further research and reproducibility.
BibTeX
@inproceedings{yilmaz2023distributed,
  author = {Selim F. Yilmaz and Can Karamanli and Deniz Gündüz},
  title = {{Distributed Deep Joint Source-Channel Coding over a Multiple Access Channel}},
  booktitle = {ICC},
  year = {2023},
  doi = {10.1109/icc45041.2023.10279499},
}
Figure