An Imperceptible and Robust Audio Watermarking Algorithm Based on SNGAN
Weili Zhou, Jiabei Zhou, Shuangyuan Yang · 2024
Digital audio watermarking has emerged as a significant research focus in the field of multimedia data hiding in recent years. However, existing deep learning-based watermarking algorithms have primarily concentrated on the image field. Moreover, current audio watermarking techniques mostly rely on traditional methods, exhibiting clear drawbacks in imperceptibility and robustness against common signal processing attacks, particularly Re-sampling and MP3 compression. To address this challenge, this paper introduces a novel digital audio watermarking approach. Leveraging the high-quality generation and outstanding generalization capabilities of Spectrally Normalization Generative Adversarial Network (SNGAN), we apply it for the first time to embed and extract watermarks in the audio domain. Additionally, we incorporate Dense Connections within the generator of the GAN to further enhance its learning capacity. Furthermore, we design a dedicated attack network tailored for common audio attacks to bolster the robustness of our model. Extensive experiments show that our proposed model achieves an advanced level of performance, with an average signal-to-noise ratio (SNR) of 26.59 and an average watermark extraction BER of 0.42%. These results conclusively establish the exceptional imperceptibility and robustness of our model.