Towards breaking DNN-based audio steganalysis with GAN

Recently, deep neural network (DNN) has significantly boosted the performance of audio steganalysis. Accordingly, most of the traditional steganography cannot resist such DNN-based steganalysis. In this work, we attempt to break a given DNN-based audio steganalysis method with a prespecified steganography. To achieve this goal, we propose to employ a generative adversarial network (GAN) to generate an enhanced cover audio firstly, which can be regarded as more suitable for steganography. Then, the secret message is embedded into the enhanced cover audio with the traditional steganography, instead of into the original cover audio. The experimental results demonstrate that our generated enhanced cover audio could effectively aid traditional steganography to break the advanced DNN-based audio steganalysis.