Generative Adversarial Analysis using U-LSB Based Audio Steganography
Vaishnavi Moorthy, Revathi Venkataraman · 2021 IEEE 18th India Council International Conference (INDICON) · 2021
Audio steganography is the technique of hiding data within a carrier audio file, where concealed data is imperceptible to humans. The existing deep-learning-based approaches depend on human handcraft for the generation and steganalysis of the steganographic audio. Generative Adversarial Network (GAN) based models are used nowadays for the generation of audio data from random latent space have proven to be efficient. This can be further utilized to strengthen the existing audio steganography methods. The proposed framework is a GAN model consisting of the U-Net-based generator, LSB-based embedder, and the discriminator. The method relies on an unsupervised adversarial training algorithm for embedding secret audio within the carrier audio in the temporal domain, making it imperceptible to humans. The experimental results on the 1-second Speech Command dataset show that the model can effectively produce steganographic audio using embedding probabilities. The steganographic method has proven to produce high-fidelity audio files capable of resisting steganalysis as compared to the audios generated using conventional methods. This technique can be effectively used for secured communication of the audio files transmission in midst of cyberspace hacking tricks.