Deep Learning-Based Adaptive Joint Source-Channel Coding using Hypernetworks

TL;DR

This paper tackles the channel adaptability challenge in deep learning–based joint source–channel coding (DJSCC) by introducing a hypernetwork-driven framework that generates channel-conditioned encoder/decoder parameters. The proposed Hyper-AJSCC employs a memory-efficient, layer-wise parameterization to map channel state ω to optimal network parameters, enabling a single model to adapt across a range of SNRs with significantly reduced storage than attention-based approaches. Empirical results on CIFAR-10 demonstrate strong adaptability for both data-oriented (image transmission) and task-oriented (image classification and cooperative inference) settings, with Hyper-AJSCC outperforming strong baselines while incurring far lower parameter overhead. The work highlights the practical impact of hypernetworks for flexible, scalable DJSCC deployments in diverse channel conditions and data tasks.

Abstract

Deep learning-based joint source-channel coding (DJSCC) is expected to be a key technique for {the} next-generation wireless networks. However, the existing DJSCC schemes still face the challenge of channel adaptability as they are typically trained under specific channel conditions. In this paper, we propose a generic framework for channel-adaptive DJSCC by utilizing hypernetworks. To tailor the hypernetwork-based framework for communication systems, we propose a memory-efficient hypernetwork parameterization and then develop a channel-adaptive DJSCC network, named Hyper-AJSCC. Compared with existing adaptive DJSCC based on the attention mechanism, Hyper-AJSCC introduces much fewer parameters and can be seamlessly combined with various existing DJSCC networks without any substantial modifications to their neural network architecture. Extensive experiments demonstrate the better adaptability to channel conditions and higher memory efficiency of Hyper-AJSCC compared with state-of-the-art baselines.

Figures (6)

• Figure 1: The considered system model of point-to-point communication with known channel condition at the transmitter and receiver.
• Figure 2: The proposed hypernetwork-based framework: A hypernetwork $H$ takes channel conditions $\omega$ as input and generates the parameters $\boldsymbol{\phi}(\omega)$ and $\boldsymbol{\theta}(\omega)$ of the encoder and decoder.
• Figure 3: The proposed hypernetwork parameterization for the $l$-th layer of DJSCC schemes. The gray panel represents the basic module, which maintains the original neural network backbone.
• Figure 4: Performance of the proposed Hyper-AJSCC compared to baseline BDJSCCs under varying training SNR with compression ratios (a) $R=1/12$ and (b) $R=1/6$. The outlined markers represent the performance of BDJSCCs when the test SNR matches their training SNR, i.e., $\text{SNR}_{\text{train}} = \text{SNR}_{\text{test}}$.
• Figure 5: Performance of the proposed Hyper-AJSCC compared to ADJSCC with different compression ratios $R$.