Adaptive MP3 Steganography Based on Adversarial Gradient

In this research, we present a new adaptive MP3 steganography technique that utilizes adversarial gradients. With the development of deep learning technology and the improvement of GPU computing performance, MP3 steganalysis methods that employ neural networks display significant detection precision under certain scenarios. To counteract these deep learning-based MP3 steganalysis methods within the realm of QMDCT (Quantization Modified Discrete Cosine Transform), we propose a new steganography method that leverages adversarial gradient. Firstly, we train a state-of-the-art deep learning steganalyzer as the adversarial model. Then, adversarial gradients are computed based on this adversarial model. Finally, these gradients are used as distortion costs that guide the embedding of data, leading the adversarial model to incorrectly classify the secret file as a benign carrier file. Our experiments demonstrate that this method is highly effective in evading detection by existing deep learning-based MP3 steganalysis methods and shows promising generalization capabilities.