Contextual Coherence Evaluation of Perfectly Secure Steganography in Text Documents

Contextual Coherence Evaluation of Perfectly Secure Steganography in Text Documents

연구 분야: Strategies

학회: International Conference on Availability, Reliability and Security

Perfectly secure steganography is a technique designed to render the presence of embedded information undetectable by ensuring that the statistical distributions before and after embedding remain perfectly identical, thereby setting the Kullback–Leibler (KL) divergence to zero. However, in terms of textual documents, specific linguistic features, such as dialects or speech patterns associated with elderly individuals, can create noticeable inconsistencies for human readers, potentially revealing the existence of hidden information. This study aims to evaluate stegotexts generated through perfectly secure steganography by employing alternative metrics beyond KL divergence. This is not a proposal to replace the KL divergence evaluation metric for perfectly secure steganography, but rather a proposal to conduct an additional evaluation using a different metric. Specifically, we utilize latent Dirichlet allocation (LDA) and cosine similarity to assess the coherence of the generated stegotext with a set of documents sharing the same topic. For our experiments, we collected news articles on gaming, sports, and politics from BBC News. The results indicate that gaming-related stegotexts exhibited significantly lower similarity with sports- and politics-related news articles, and a similar trend was observed for politics-related stegotexts. However, for sports-related stegotexts, no significant difference was detected.