WavInWav: Time-domain Speech Hiding via Invertible Neural Network

TL;DR

WavInWav presents a time-domain audio hiding framework based on a flow-based invertible neural network (WavINN) that directly conceals a complete secret audio within a cover audio. The method enforces time-frequency consistency through a multi-resolution STFT loss, incorporates a composite differentiable noise layer for robustness, and adds an encryption/decryption module to protect hidden data, enabling end-to-end concealment and recovery via inverse network passes. Experiments on LibriSpeech and VCTK show superior objective and subjective performance for both stego and recovered secret audio, with notable SNR gains up to $\approx 19.5$ dB in secret recovery and strong generalization to unseen speakers/datasets. The approach demonstrates practical potential for secure audio transmission and targeted communications, while acknowledging limitations in strong-noise resistance and steganalysis defenses, and outlining avenues for extending to other audio formats and domains.

Abstract

Data hiding is essential for secure communication across digital media, and recent advances in Deep Neural Networks (DNNs) provide enhanced methods for embedding secret information effectively. However, previous audio hiding methods often result in unsatisfactory quality when recovering secret audio, due to their inherent limitations in the modeling of time-frequency relationships. In this paper, we explore these limitations and introduce a new DNN-based approach. We use a flow-based invertible neural network to establish a direct link between stego audio, cover audio, and secret audio, enhancing the reversibility of embedding and extracting messages. To address common issues from time-frequency transformations that degrade secret audio quality during recovery, we implement a time-frequency loss on the time-domain signal. This approach not only retains the benefits of time-frequency constraints but also enhances the reversibility of message recovery, which is vital for practical applications. We also add an encryption technique to protect the hidden data from unauthorized access. Experimental results on the VCTK and LibriSpeech datasets demonstrate that our method outperforms previous approaches in terms of subjective and objective metrics and exhibits robustness to various types of noise, suggesting its utility in targeted secure communication scenarios.

Figures (4)

• Figure 1: Comparison between framework relying on spectrums and our flow-based framework with time-domain hiding.
• Figure 2: Overview of the proposed audio hiding method. The audio is hidden in the time domain, while loss is computed in the time-frequency domain.
• Figure 3: The structure of the invertible neural network and noise layers used in our approach. The forward and backward processes share the same parameters.
• Figure 4: The backbone of the $i$-th invertible block used in our approach.