CsiMix-JSCC: A Lightweight Adaptive Deep Joint Source Channel Coding Framework for CSI Feedback

Abstract

In frequency-division-duplexing (FDD) massive multiple-input multiple-output (mMIMO) systems, accurate downlink channel state information (CSI) feedback is essential for achieving high spectral efficiency. However, the high dimensionality of CSI leads to substantial uplink feedback overhead and makes the feedback process sensitive to channel impairments. Existing deep joint source-channel coding (DJSCC)-based CSI feedback methods still face two practical limitations: excessive encoder complexity at resource-constrained user equipment (UE) and limited adaptivity to varying uplink channel conditions. This paper proposes CsiMix-JSCC, a lightweight adaptive DJSCC framework for CSI feedback. It employs a mixer-based autoencoder with per-antenna tokenization to reduce UE-side complexity and incorporates signal-to-noise ratio (SNR)-aware conditioning to enable a single model to operate across varying uplink conditions. Numerical results on the COST2100 dataset show that CsiMix-JSCC reduces the normalized mean-square error (NMSE) by up to 44.5% relative to the adaptive DJSCC baseline ADJSCC and by up to 64.3% relative to the DJSCC-CSINet baseline at $\mathbf{S N R} \boldsymbol{=} \mathbf{1 0 ~ d B}$. It also achieves average NMSE reductions of 32.8% and 51.7% over SNR in [-10, 10] dB, respectively. Meanwhile, compared with ADJSCC, it reduces UEside FLOPs by 95.6% and parameters by 65.6%, demonstrating a favorable tradeoff among robustness, reconstruction accuracy, and implementation efficiency.