Encrypted rich-data steganography using generative adversarial networks

Abstract

Steganography has received a great deal of attention within the information security domain due to its potential utility in ensuring network security and privacy. Leveraging advancements in deep neural networks, the state-of-the-art steganography models are capable of encoding a message within a cover image and producing a visually indistinguishable encoded image from which the decoder can recover the original message. While a message of different data types can be converted to a binary message before encoding into a cover image, this work explores the ability of neural network models to encode data types of different modalities. We propose the ERS-GAN (Encrypted Rich-data Steganography Generative Adversarial Network) - an end-to-end generative adversarial network model for efficient data encoding and decoding. Our proposed model is capable of encoding message of multiple types, e.g., text, audio and image, and is able to hide message deeply into a cover image without being detected and decoded by a third-party adversary who is not given permission to access the message. Experiments conducted on the datasets MS-COCO and Speech Commands show that our model out-performs or equally matches the state-of-the-arts in several aspects of steganography performance. Our proposed ERS-GAN can be potentially used to protect the wireless communication against malicious activity such as eavesdropping.