Low-Complexity Detection of CIM-SSK over Rayleigh MIMO Channels: A Deep Learning Approach
2026 Joint European Conference on Networks and Communications and 6G Summit, EuCNC/6G Summit 2026, Malaga, İspanya, 2 - 05 Haziran 2026, ss.1034-1039, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/eucnc/6gsummit68295.2026.11577517
- Basıldığı Şehir: Malaga
- Basıldığı Ülke: İspanya
- Sayfa Sayıları: ss.1034-1039
- Anahtar Kelimeler: CIM-SSK, Code index modulation, convolutional neural networks, deep learning, Hadamard codes, list-based detection, MIMO, Rayleigh block fading, residual networks, space shift keying
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
Code index modulation (CIM) combined with space shift keying (SSK), referred to as CIM-SSK, increases spectral efficiency by embedding information jointly into the active transmit-antenna index and the spreading-code indices assigned to the in-phase and quadrature components. However, conventional exhaustive joint maximum-likelihood (ML) detection becomes increasingly burdensome as the antenna and codebook sizes grow. In this study, a deep-learning (DL)-assisted receiver is presented for CIM-SSK over Rayleigh block-fading multiple-input multiple-output (MIMO) channels with additive white Gaussian noise (AWGN). Receiver-side inputs are formed from channel state information, correlator-bank statistics, raw chip-rate observations, and explicit noise/signal-to-noise ratio (SNR) side information, which are organized into a structured three-channel tensor that preserves the antenna/code/chip ordering. A compact residual CNN2D is then used to produce a probabilistic ranking of the joint hypotheses, and a short Top-K candidate list is generated. The final decision is then obtained by applying the chip-domain Euclidean refinement only to the Top-K candidates. Bit error rate (BER) results indicate gains relative to the considered CIM-SSK joint-ML baseline as well as equal-rate SSK and SM benchmarks, while the CNN ranking stage achieves 95.92% Top-1 and 98.99% Top-4 accuracy for 64 hypotheses, supporting reliable shortlist-based detection with reduced online search.