PixInWav: Hiding Pixels in music with Deep Steganography - Margarita Geleta - UPC Barcelona 2020

Project page (updated after this video): https://imatge.upc.edu/web/publications/pixinwav-residual-steganography-hiding-pixels-audio Presentation in the Challenge Based Innovation (CBI) module of the Datas Science and Engineering Bachelor program at UPC TelecomBCN. Steganography comprises the mechanics of hiding secret data within a cover media which may be publicly available with the main premise that the fact that the communication is taking place is hidden as well. In this research work, classic methods for audio steganography are studied, followed by the proposal of a new audio steganography model with deep neural networks which embeds image data into audio. Moreover, unlike many popular audio models, the system includes a short-time Cosine transform and inverse-short-time Cosine transform to change from time domain to frequency and back. This approach distributes the spectral changes as well as the short-time Fourier transform and has almost no reconstruction loss. Lastly, an evaluation of the model has been carried out under different architecture alterations. Qualitative experiments suggest that the decoded images are highly identifiable.