NAS-StegNet: Lightweight Image Steganography Networks via Neural Architecture Search

Deep steganography describes the task of hiding a full image in another for secret communication, and such a model usually consists of hide (H) network for secret hiding followed by revealing (R) network for secret revealing. To guarantee the hiding effect for the secret communication applications (e.g., watermarking and light field messaging), most of existing deep steganography models design complex network architecture for H and R, increasing the challenge for model deployment. To achieve a better trade-off between steganography effect and model complexity, in this paper, we explore the idea of neural architecture search to learn a more practical deep steganography network, which is able to produce powerful steganography results but with much less parameters. Specifically, our automatically-learned network, termed as NAS-StegaNet, has 26 $$\times $$ fewer parameters and requires 2 $$\times $$ fewer GFLOPs when compared with the most powerful model. Codes are available at https://github.com/wang-MIG-CFM-UESTC/nas_stegan.git .