Transfer Learning Based Detection for Intelligent Reflecting Surface Aided Communications

This work investigates the data detection problem in an Intelligent Reflecting Surface (IRS) aided downlink communication between a multi-antenna access point (AP) and multiple user equipments (UEs). We utilise a deep learning-based approach, with a maximum likelihood detection (MLD)-based loss function, thereby bypassing the resource-consuming channel training and estimation requirement for detection. The proposed detection framework first trains a base deep neural network (DNN) offline with the simulated samples of the channel coefficients and IRS phase shifts in the IRS-assisted communications scenario. To deal with the significant challenge of the channel getting outdated, domain adaptation under the transfer learning paradigm is leveraged, i.e., the initial layers of the DNN are frozen, and the remaining layers are retrained on a smaller number of the received signal samples online to account for the channel mismatch. Our results show that the proposed detector achieves BER results close to the lower bound and outperforms conventional benchmark techniques, with relatively lower complexity.