Knowledge Base Identification and Adaptation for End-to-End Semantic Communication

As a critical technology for sixth-generation (6G) communication systems, semantic communication has attracted significant attention due to its high communication efficiency. However, when there is a knowledge base mismatch between the transmitter and the receiver in semantic communication systems, the communication performance may degrade significantly. Therefore, we propose a feature-based identification scheme to assess the degree of match between the transmitter's and receiver's semantic knowledge bases (SKBs). Specifically, we design a feature vector with the potential to replace the embedding matrix for characterizing similarity across different knowledge bases. Moreover, to mitigate the performance degradation under SKB mismatch, we propose a lightweight, intermediate-layer-based SKB adaptation scheme, accompanied by two specifically designed training algorithms. Simulation results demonstrate that the proposed SKB identification scheme can effectively identify SKBs in most cases while significantly reducing the required transmission load. The proposed SKB adaptation scheme can significantly enhance communication performance in SKB-mismatched scenarios with reasonably low additional computational and storage overhead.