A Lightweight Low-Complexity CSI Feedback Network for 6G Massive MIMO Systems

Abstract

Massive multiple-input multiple-output (MIMO) systems require accurate channel state information (CSI) at the base station, whereas frequency-division-duplex downlink CSI feedback introduces substantial uplink overhead. Autoencoder-based CSI feedback networks reduce feedback dimensionality, but their parameter size and reconstruction cost may still be non-negligible for lightweight 6G terminals. This paper proposes LightCSI-Net, a compact CSI feedback network that integrates depthwise separable convolution, squeeze-and-excitation channel attention, latent compression, and residual reconstruction. Experiments are conducted on indoor and outdoor COST2100-format small-scale channel data under compression ratios of 1/4, 1/8, 1/16, and 1/32. The evaluation uses normalized mean square error (NMSE), cosine similarity, trainable parameters, and single-sample inference time. At the 1/16 compression ratio, LightCSI-Net reduces trainable parameters by approximately 49.3% compared with CsiNet while maintaining comparable indoor NMSE and cosine similarity. However, its measured CPU inference time increases from 0.091 ms to 0.283 ms, showing that parameter reduction does not automatically imply lower latency. We therefore interpret the proposed model as parameter-and memory-efficient, and provide a hardware-aware latency discussion for URLLC-oriented deployment.