Steganography has been largely studied in the computer vision community for digital watermarking or secret messaging. Recently, various deep learning-based approaches have been proposed to tackle the steganography task. However, in terms of capacity, existing methods are not enough to hide large data types like video content due to its high bitrate. In this thesis, we push this limit by investigating methods for hiding video content inside audio files. To solve this novel cross-modal steganography, we first introduce several models based on existing methods as the baseline. However, we discovered the clear limitations of those models in evaluation. Finally, to mitigate these limitations, we devise a new optimization-based method named HvFO to improve not only the reconstructed v...[ Read more ]

Steganography has been largely studied in the computer vision community for digital watermarking or secret messaging. Recently, various deep learning-based approaches have been proposed to tackle the steganography task. However, in terms of capacity, existing methods are not enough to hide large data types like video content due to its high bitrate. In this thesis, we push this limit by investigating methods for hiding video content inside audio files. To solve this novel cross-modal steganography, we first introduce several models based on existing methods as the baseline. However, we discovered the clear limitations of those models in evaluation. Finally, to mitigate these limitations, we devise a new optimization-based method named HvFO to improve not only the reconstructed video quality but also embedded audio fidelity.

The HvFO uses recent advances in flow-based generative models that enable effective mapping audio to latent codes with further optimization so that nearby codes correspond to perceptually similar signals. We show that compressed video data can be concealed in the latent codes of audio sequences while preserving the fidelity of the hidden video and the original audio. We can embed 128x128 video inside same-duration audio, or higher-resolution video inside longer audio sequences. Quantitative experiments show that our approach outperforms relevant baselines in steganographic capacity and fidelity.