SCAN: Semantic Communication With Adaptive Channel Feedback

Guangyi Zhang, Qiyu Hu, Yunlong Cai, Guanding YuIEEE Trans. Cogn. Commun. Netw 2024 · DOI

TL;DR

This paper introduces SCAN, a novel framework for semantic communication that improves image transmission reliability by adaptively adjusting channel feedback based on predicted image reconstruction quality. By introducing a new metric, Semantic Distortion Outage Probability (SDOP), and a performance evaluator, SCAN significantly reduces feedback overhead while ensuring high-quality image reconstruction.

What Problem Does This Paper Address?

Existing semantic communication systems, while promising for efficient data transmission, often struggle with inconsistent image reconstruction quality. This variability leads to uncertainty in reliable transmission, especially when channel conditions are poor. The paper highlights that current systems focus on average distortion (like PSNR) and fail to account for the varying robustness of different images to disturbances. Images requiring higher reconstruction quality are more susceptible to errors, necessitating a more adaptive approach to resource allocation, particularly for channel feedback, which can be a significant overhead in MIMO systems.

Methods

The core of the proposed solution is the Semantic Communication with Adaptive channel feedback (SCAN) framework. Key components include:

  • DeepSC-MIMO: A deep learning-based semantic communication system designed for MIMO channels. It leverages channel state information (CSI) and noise variance to adaptively allocate power to important feature vectors through an attention mechanism.
  • Semantic Distortion Outage Probability (SDOP): A new performance metric introduced to quantify the reliability of semantic communication systems. SDOP is defined as the probability that the instantaneous distortion exceeds a given target threshold.
  • Performance Evaluator: A lightweight deep neural network trained using knowledge distillation. This evaluator predicts the reconstruction quality of an image at the transmitter before transmission, allowing for proactive resource allocation.
  • Adaptive Channel Feedback: SCAN employs two adaptive strategies:
  • Instance-wise: Adjusts CSI compression level for each image individually based on its predicted reconstruction quality. Images predicted to have lower quality receive more feedback bits.
  • Group-wise: Optimizes CSI compression levels for a group of images under an average feedback length constraint, aiming to minimize the overall SDOP for the group.

Datasets

  • CIFAR10 dataset: Used for training and testing the DeepSC-MIMO model. It consists of 50,000 color images (32x32x3) for training and 10,000 for testing.

Key Insights and Ideas

  • Image-specific reliability is crucial: The paper emphasizes that not all images are created equal in semantic communication. Their inherent complexity and the required reconstruction quality significantly impact their robustness, necessitating an image-aware approach.
  • SDOP as a better reliability metric: Traditional metrics like average PSNR don't capture the "worst-case" performance, which is critical for reliability. SDOP provides a more direct measure of transmission failure probability.
  • Predicting quality enables adaptation: The ability to predict an image's reconstruction quality before transmission is the linchpin of SCAN. This prediction allows for intelligent CSI compression, allocating more resources to images that need them most.
  • Knowledge distillation for efficiency: Using knowledge distillation to train the performance evaluator makes it lightweight and efficient, allowing it to run at the transmitter without significant computational burden.

Key Results

  • Improved Reliability: SCAN significantly improves SDOP performance, meaning fewer images are reconstructed with distortion above the target threshold.
  • Reduced Feedback Overhead: The adaptive nature of SCAN leads to substantial reductions in channel feedback overhead compared to fixed-length feedback schemes. For instance, Figure 11(b) shows a significant reduction in average required length for a given SDOP.
  • Accurate Performance Prediction: The performance evaluator demonstrates high accuracy in predicting PSNR values, with relatively low prediction error across various SNR and CSI codeword lengths (Figure 9).
  • Competitive Performance: DeepSC-MIMO, the underlying semantic communication system, outperforms traditional DJSCC and BPG+LDPC schemes, especially in low SNR regimes, by effectively utilizing CSI and noise information.

Limitations and Future Work

  • Generalization to other modalities: The authors suggest that SCAN is a general framework and future work could extend it to other data modalities beyond images.
  • Further adaptive designs: The paper points out that within the SCAN framework, more adaptive designs, such as adaptive coding and adaptive modulation, could be explored for even higher performance gains.
  • Computational complexity of training: While the performance evaluator is lightweight, the training of the DeepSC-MIMO model itself can be computationally intensive. The adaptive design in SCAN-A, however, is shown to reduce training time.

Summary updated

BibTeX
@article{zhang2024scan,
  author = {Guangyi Zhang and Qiyu Hu and Yunlong Cai and Guanding Yu},
  title = {{SCAN: Semantic Communication With Adaptive Channel Feedback}},
  journal = {IEEE Trans. Cogn. Commun. Netw},
  year = {2024},
  doi = {10.1109/tccn.2024.3394867},
}
Figure