Audio steganography cover enhancement via reinforcement learning

Recent advancements in steganography analysis based on deep neural networks have led to the development of steganography schemes that incorporate deep network technology like adversarial example, GAN, and reinforce learning. However, most deep network-based steganography schemes are unable to ensure error-free extraction of secret information because of the design similar to information reconstruction. To address this issue, this work proposes a novel audio steganography cover enhancement framework that leverages two networks—a policy network and an environment network—to enhance the undetectability and imperceptibility of the steganographic audio. The proposed schemes utilize the reinforcement algorithm to optimize steganography cover modification for improving undetectability and imperceptibility. And our method guarantees 100% extracting accuracy by only enhancing on domains that do not affect information extraction. The experimental results demonstrate that our method significantly enhances the missing detection rate of the target audio steganography analysis network over 90% when resisting a steganalysis network, and meanwhile, our method has strong imperceptibility which can hardly be distinguished by human ears.