Variable Length Variable Quality Audio Steganography

Seungmo Ku

Keywords: computer vision, stegaography, recurrent neural network, loss conditional training, information hiding

Abstract: Steganography is the task of hiding and recovering secret data inside a non-secret container data while making imperceptible changes to the container. When using steganography to hide audio inside an image, current approaches neither allow the encoding of a signal with variable length nor allow making a trade-off between secret data reconstruction quality and imperceptibility in the changes made to the container image. To address this problem, we propose VLVQ (Variable Length Variable Quality Audio Steganography), a deep learning based steganographic framework capable of hiding variable-length audio inside an image by training the network to iteratively encode and decode the audio data from the container image. Complementary to the standard reconstruction loss, we propose an optional conditional loss term that allows the users to make quality trade-offs between audio and image reconstruction on inference time, without needing to train a separate model for each trade-off setups. Our experiments on ImageNet and AudioSet demonstrate VLVQ’s ability to retain reasonable image quality (28.99 $psnr$) and audio reconstruction quality (23.79 $snrseg$) while encoding 19 seconds of audio. We also show VLVQ’s capability to generalize to signals longer than what is seen during training.

One-sentence Summary: We propose a steganographic system that can hide variable-length audio signals inside an image and can make inference time quality tradeoffs.