A Secure Data Hiding Approach Using Natural Language Processing

Abstract Steganography involves the creation of stegosystems that conceal data within a cover medium. In this work, we propose a linguistic stegosystem that employs natural language as the concealment medium using a Word2Vec-based continuous bag-of-words (CBOW) model with noise contrastive estimation. The system selects from the top 8 predicted candidate words at each embeddable position, achieving an embedding capacity of up to 3 bits per word (bpw) and a practical embedding utilization of 44.83% (up to 50% maximum capacity) of eligible cover words. Experimental evaluation demonstrates a perplexity of 10.3, outperforming several existing linguistic steganography approaches while maintaining high semantic coherence. Steganalysis experiments show limited detectability, with detection accuracies of 58% using TS-RNN, 64% using a BERT-based classifier (AUC = 0.70), and a noise-based perturbation flag rate of 12%, indicating strong resistance against statistical and neural detection methods. The proposed approach balances embedding capacity, linguistic naturalness, and security while maintaining computational efficiency through the lightweight CBOW architecture.