Enhancing Audio Steganography Through CGAN-Generated Cover Audio and Adaptive LSB Embedding: A Hybrid Approach
Audio steganography hides secret messages inside audio files, enabling covert communication without drawing attention. Audio steganography methods aim to achieve high imperceptibility, robust performance, and high payload capacity. While traditional techniques like Least Significant Bit (LSB) coding offer good imperceptibility, they are highly vulnerable to statistical steganalysis and signal manipulation. Existing hybrid methods suffer from maintaining quality across diverse audio and inadequate robustness mechanisms, struggling to balance imperceptibility, payload capacity, and robustness. This paper proposes a novel hybrid approach that combines Conditional Generative Adversarial Networks (CGANs) with LSB coding to address these limitations. The CGAN is trained with the LibriSpeech dataset to generate audio patterns that simulate spontaneous speech for use as adaptive covers. The model was implemented using PyTorch, with performance evaluated based on Signal-to-Noise Ratio (SNR), Perceptual Evaluation of Speech Quality (PESQ), Bit Error Rate (BER), and robustness to audio transformations. Experimental results showed a PESQ value of 4.05 and a mean SNR of 33.1 dB, representing excellent audio quality given the substantial payload capacity of 1.27 kbps. The method achieved a BER value of 2.23% and 87% robustness to compression, filtering, and resampling operations. The effectiveness of the CGAN-LSB hybrid method for enhancing capacity, imperceptibility, and robustness is achieved through the CGAN's ability to generate statistically natural audio covers, while adaptive LSB integration preserves data integrity during signal processing operations, making it highly suitable for secure audio communication and audio watermarking applications. While the generated covers exhibit distributional properties similar to genuine audio, direct validation against specific steganalysis detectors remains an important direction for future empirical evaluation.
