DQN-Based Dynamic Private Knowledge Graph Construction for Secure Semantic Communication

Abstract

Semantic communication (SemCom) has emerged as a transformative technology for 6G networks, enabling ultra-efficient data transmission through context-aware encoding. However, it also raises new privacy challenges, as eavesdroppers with advanced semantic models can infer sensitive information from minimal intercepted data. This paper proposes PrunedCBA-DQN, a privacy-preserving SemCom framework that constructs private knowledge by extracting subgraphs from the sender’s knowledge graph using source messages as keywords. This private knowledge is integrated into the encoder-decoder pipeline, ensuring that critical semantics remain inaccessible without it. We leverage the Diameter-bounded max-Coverage Group Steiner Tree (DCGST) formulation for expressive keyword-based graph exploration and train deep reinforcement learning (DRL) agents to optimize the diameter parameter dynamically. Simulations show that our approach achieves high-quality semantic reconstruction while effectively countering eavesdropping, advancing secure and intelligent 6G SemCom systems.