Model Shift Modulation for Semantic Communication

In recent years, semantic communication (SemCom) has garnered considerable attention within communication fields. This study presents a great attempt to exploit the use of semantic feature shifts in images to enable the joint transmission of semantic features and precise digital signals. This method enhances SemCom by facilitating the transmission of digital signals with relatively low bit error rates without making major modification to the SemCom. Inspired by the intrinsic semantic features of images and the relationship between the semantic feature and the original image, this paper propose Model Shift Modulation (MSM) for SemCom. Specifically, the digital signal will be converted into a binary stream and mapped to distinct shifts in the image’s feature map, resulting in the pixel shifts of the images that do not degrade the quality of image reconstruction. For demodulation at the receiver, deep learning is used to perform edge detection on the decoded shifted images to achieve the demodulation task of recovering the image and binary stream. Experimental results show that MSM can reliably transmit digital signal while ensuring good image reconstruction performance under specific signal-to-noise ratios (SNRs).