Audio watermarking is an important technique for copyright protection, multimedia authentication, and secure content distribution. However, many existing methods still struggle to achieve a reliable balance among imperceptibility, robustness, and embedding capacity, mainly because watermark insertion is often performed using fixed rules that do not sufficiently reflect the local behaviour of audio signals. This paper proposes an adaptive audio watermarking framework in which each processing stage is designed to support a specific decision in the embedding and extraction pipeline. First, empirical mode decomposition (EMD) is used to decompose each audio frame into intrinsic mode functions (IMFs), providing a signal-adaptive representation of the non-stationary host audio. Then, a 1D convolutional neural network(1D-CNN) extracts representative features from these components. Based on these features, K-Means++ clustering identifies stable and perceptually suitable IMFs with favourable energy and variance characteristics. The watermark is embedded in the SVD domain by modifying the dominant singular values through a 2-bit quantization strategy, which improves payload capacity while preserving audio quality. Finally, an XGBoost classifier learns the selected embedding locations and supports blind watermark extraction. Experiments on four audio genres show that the proposed method achieves an average SNR of 45.1 dB, ODG of −0.28, embedding capacity of 2350 bps, and perfect extraction under no-attack conditions with BER = 0 and NC = 1.0. The method also maintains low BER and high NC under Stirmark and common signal-processing attacks, making it suitable for secure audio distribution and copyright protection.
Mohialden et al. (Fri,) studied this question.