StyleCloak: Anonymous Source Coding for Textual Attribute Obfuscation with Semantic Fidelity

With the explosion of ubiquitous user information, privacy preservation has emerged as a critical challenge in next-generation real time communications, spurring research of anonymous communications. Existing anonymity designs primarily focus on masking user attributes hidden in transmitted signals, yet user characteristics embedded in data sources can still be exploited by intelligent adversaries empowered by large language models (LLMs). In this pioneers work, we investigate anonymous source coding design, specifically for textual data sources. A novel anonymous framework, namely StyleCloak, is proposed to obfuscate user-specified attributes through minimal yet effective alterations to lexical features. By proposing a novel Pareto Beam search and iteration, StyleCloak achieves high obfuscation efficiency while well preserving semantic meaning. In addition, StyleCloak’s plug-and-play architecture enables flexible adaptation to various user attribute anonymization tasks. It can be integrated as a preprocessing module for diverse downstream applications, such as real time semantic communication, without modifying existing system architectures.