METHOD OF STEGANOGRAPHIC HIDING IN AUDIO CONTAINERS USING LLMS

Abstract

Background. The development of digital networks requires the improvement of methods for covert data transmission, among which the LSB approach is the most common due to its simplicity and high capacity. However, modern stegoanalysis methods necessitate a deep systematization and critical analysis of existing research to increase the stability and adaptability of algorithms. The goal is a comprehensive analysis and systematization of modern research in the field of LSB steganography to identify current trends and form an analytical base that will allow improving methods for covert information transmission. Methods. The method of systematic analysis and classification of modern scientific publications in the field of digital data protection was used. The study covered the analysis of the evolution of LSB (Least Significant Bit) methods, in particular their hybridization with cryptographic algorithms (AES, RSA), frequency transforms (DWT, DCT) and stochastic approaches (chaotic maps). Special attention was paid to the use of Reed-Solomon coding, adaptive pixel selection and machine learning models to optimize data embedding. Results. An adaptive audio steganography method has been developed that uses large language models (LLM) to dynamically control the depth of LSB embedding ($k$) based on psychoacoustic analysis and local signal entropy. Experimental simulation in MATLAB confirmed the superiority of the method over classical analogues, in particular, the highest peak signal-to-noise ratio (PSNR = 58.23 dB) and the minimum mean square error (MSE = 0.098) were achieved. Conclusions. The integration of LLM into the steganographic coding process allows us to transform the LSB approach into an intelligent system that adapts to the semantic context of the audio container, providing high visual and statistical invisibility. Despite the increase in processing time, the proposed algorithm is optimal for secure data transmission systems due to its ability to minimize the impact on the human auditory system (HAS) and resist modern stegoanalysis methods.