Multimodal Large Language Models Driven Privacy-Preserving Wireless Semantic Communication in 6G

In the context of 6G and artificial intelligence, the amount of data on the internet is rapidly increasing. However, traditional syntactic communication encoding techniques are approaching the Shannon limit and have reached a bottleneck. Semantic communication can significantly reduce data traffic and enhance communication efficiency by converting multimodal data, such as images, audio, and video, into semantic information for transmission. However, current semantic communication technology is still in its early stages and faces many challenges. The main challenges in semantic communication currently include data heterogeneity, semantic consistency, and privacy issues. Multimodal large language models (MLLM) could perform semantic understanding of multimodal data and semantic-based multimodal data generation, which offer a unified and consistent multimodal semantic conversion. Therefore, we propose a Privacy-Preserving Semantic Communication scheme based on MLLM (MLLM-PSC). The semantic interpretation ability of MLLM-PSC is cultivated based on the pre-training and fine-tuning of MLLM, without the need for additional knowledge base (KB) alignment. We take textual semantics as a medium for consistently conversion of multimodal data, while safeguarding sensitive semantic information based on few-shot learning. Our simulation results demonstrate the superior performance of MLLM-PSC.