Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2509.04005

Robust MIMO Semantic Communication with Imperfect CSI via Knowledge Distillation

Mingze Gong·Shuoyao Wang·Shijian Gao·Jia Yan·Suzhi Bi
View original source ↗Read source document ↗Export metadata ↓

Abstract

Semantic communication (SemComm) has emerged as a new communication paradigm. To enhance efficiency, multiple-input-multiple-output (MIMO) technology has been further integrated into SemComm systems. However, existing MIMO SemComm systems assume perfect channel matrix estimation for channel-adaptive joint source-channel coding, which is impractical due to hardware and pilot overhead constraints. In this paper, we propose a semantic image transmission system with channel matrix and channel noise adaptation, named HANA-JSCC, to cope with channel estimation errors in MIMO systems. We propose a channel matrix adaptor that collaborates with the channel codec to adapt to misaligned channel state information, thereby mitigating the impact of estimation errors. Since the relationship between the estimated channel matrix and true channel matrix is ill-posed (one-to-many), we further introduce a two-stage training strategy with knowledge distillation to overcome the convergence difficulties caused by the ill-posed problem. Comparing with the state-of-the-art benchmarks, HANA-JSCC achieves $0.40\sim0.54$dB higher average performance across various noise and estimation error levels in various datasets.

Explore related subjects

eess.SP

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Mingze Gong, Shuoyao Wang, Shijian Gao, Jia Yan, Suzhi Bi. 2025-09-04. Robust MIMO Semantic Communication with Imperfect CSI via Knowledge Distillation. https://arxiv.org/abs/2509.04005

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Joint Detection and Velocity Estimation in OFDM-ISAC Cell-Free Massive MIMO Networks

This paper develops a sensing framework for cell-free massive MIMO (CF-mMIMO) networks operating under orthogonal frequency division multiplexing (OFDM)-based integrated sensing and communication (ISAC). The framework explicitly incorporates the 3D bistatic Doppler geometry across distributed access points (APs) into a generalized likelihood ratio test (GLRT) detector. To address system scalability, a user-target-centric AP association approach is utilized. The 3D velocity vector of the target is estimated, and several search and optimization strategies, including coarse grid search, gradient-based refinement, and particle swarm optimization (PSO), are developed and evaluated. Simulation results demonstrate that the proposed PSO-aided detector achieves the most favorable accuracy-complexity trade-off, while Doppler mismatch can cause substantial sensing signal-to-noise ratio (SNR) degradation in high-mobility scenarios. Additionally, leveraging more OFDM subcarriers enhances frequency-domain diversity and yields further sensing-SNR gains.

eess.SP↗
arXiv2026

Mixture-of-Experts Transformer for Automatic Modulation Recognition

Automatic Modulation Recognition (AMR) is a key enabling technology for cognitive radio and intelligent spectrum management in next-generation wireless systems. However, current deep learning-based AMR methods predominantly rely on static multi-scale fusion strategies, which lack the flexibility to adapt to the highly dynamic temporal variations of modulation signals. To address this limitation, we propose MoEformer, an adaptive Multi-Scale Mixture-of-Experts Transformer network that directly processes I/Q signals to preserve their temporal and phase structures. Specifically, MoEformer constructs multi scale expert views through temporal resampling, employs an input-dependent gating mechanism for dynamic expert fusion, and integrates Rotary Position Embeddings (RoPE) within Transformer encoders to capture both local and global tem poral dependencies. Comprehensive evaluations on three widely adopted benchmarks (RadioML2016.10a, RadioML2016.10b, and RadioML2018.01A) demonstrate that MoEformer outperforms the competitive baselines, achieving superior average recognition accuracies of 63.74%, 66.24%, and 64.22%, respectively. In addition, the proposed method strikes an optimal trade-off between recognition performance and model complexity.

eess.SP↗
arXiv2026

Hierarchical Federated Learning for Unsupervised Waveform Classification over Tactical MANETs

Distributed radio frequency sensing in contested tactical environments demands collaborative learning across mobile nodes. In ad-hoc networks, learning must occur without persistent backhaul, ground truth labels, or reliable communication links. Traditional federated learning approaches assume either ideal link conditions or supervised training objectives, neither of which holds in practice for deployed MANET platforms. This paper presents a hierarchical federated learning framework for unsupervised waveform classification over tactical MANETs subject to Rayleigh fading, random waypoint mobility, and multi-hop routing loss. Each node trains a local denoising convolutional autoencoder on raw IQ observations without label exchange, learning compact representations through a self-supervised reconstruction objective. A two-stage aggregation protocol elects connectivity-based relay aggregators consistent with OLSR multipoint relay selection, compressing cluster-level model updates before forwarding to a mobile server proxy. Across five random seeds, in-network aggregation reduces attempted transmission bits by 16% on average relative to relay-forward federated averaging at comparable classification performance. Despite mean per-round update drop rates of 21% (hierarchical) and 34% (flat), hierarchical MANET FL attains the highest mean unsupervised representation quality of the three federated conditions tested while also showing substantially lower run-to-run variance than either flat MANET FedAvg or ideal FedAvg; flat routing itself shows no consistent benefit or penalty relative to ideal FedAvg. Performance is assessed using KMeans normalized mutual information and linear probe accuracy on the learned latent embeddings.

eess.SP↗