Covert Text Communication Using GPT-Based Language Models and Semantic Similarity Constraints
Abstract
Covert text communication using GPT-based language models and semantic similarity constraints aims to conceal messages within natural-sounding text while maintaining the intended meaning. This method leverages the generative capabilities of large language models, such as GPT, to produce linguistically coherent outputs that embed secret messages. Existing covert communication techniques often suffer from delectability due to unnatural phrasing, limited semantic control, or vulnerability to steganalysis. To address these challenges, we propose a framework that combines GPT-based text generation (GPT-TG) with the embedding of hidden content. In this approach, secret messages are encoded into prompts that guide GPT to generate contextually appropriate text, and a semantic similarity constraint ensures message fidelity and stealth. The proposed method enables secure, human-like covert communication suitable for low-bandwidth, text-based channels. Experimental results demonstrate that the generated covert texts are indistinguishable from normal language, achieving high semantic preservation and low detection rates using standard steganalysis tools.