Knowledge Discrepancy Oriented Privacy Preserving for Semantic Communication

Semantic communication is an emerging paradigm relying on deep neural network alongside knowledge bases to enable correct delivery of desired information meaning. Most existing semantic communication systems assume that communication nodes share the same public knowledge base, while ignoring the agnostic and diversity of private knowledge that may lead to a knowledge discrepancy. Consequently, the receiver can infer sensitive information that the sender does not want to disclose. To address this privacy risk, we provide a knowledge discrepancy oriented privacy preserving method (KDPP) for semantic communication. KDPP aims to narrow down the discrepancy between prior knowledge and receiver's knowledge before privacy preserving. Specifically, we propose knowledge mapping to solve the agnostic problem, matching the unknown knowledge to the known prior knowledge. Moreover, we introduce disambiguation to address the diversity issue, capturing diverse interpretations of ambiguous concepts. Our simulations show that KDPP performs better than existing methods in terms of the inference accuracy and the privacy preserving effect. This demonstrates the validity of private knowledge inference as an effective method for privacy preserving against knowledge discrepancy.