Autoencoder-based OFDM-IM with joint index-symbol classification via attention mechanisms

The performance of orthogonal frequency division multiplexing with index modulation degrades under realistic channel conditions due to suboptimal detection and decoding strategies. This paper presents a deep learning framework using a structured, end-to-end autoencoder for joint index and symbol detection. The model integrates convolutional, multi-head attention, and dense layers to extract features and incorporate channel knowledge via minimum mean square error equalization. The attention mechanism enables the network to focus on critical features of the received signal and channel state information. In conventional systems, the lack of task-specific supervision for index and symbol detection leads to continuous-valued outputs that are not easily mappable to discrete transmitted bits, degrading data recovery. To address this, a classification-based loss function is designed to recover both active subcarrier indices and symbols. Extensive experiments demonstrate robustness under Rayleigh fading channels, with significant improvements in block error rate and bit error rate compared to conventional techniques. • A novel end-to-end AE-based OFDM-IM framework is proposed for joint transmitter–receiver optimization via gradient-based learning. • A task-specific multi-branch decoder with attention is designed to separately detect subcarrier indices and modulation symbols. • CSI is integrated via MMSE equalization and explicitly injected into the neural decoder for robust detection under Rayleigh fading. • A joint multi-task loss improves BER and BLER performance under perfect and imperfect CSI across a wide SNR range.