Multi - Domain HEVC video steganography based on invertible neural network

Abstract Most existing research on High Efficiency Video Coding (HEVC) video steganography is based on the premise that the receiver can acquire an unaltered bitstream. However, videos uploaded to social media platforms (e.g., YouTube, WeChat, TikTok) are prone to be re-encoded. This process can corrupt or entirely remove the information hidden in the original bitstream. To address this issue, this paper proposes a novel scheme that jointly leverages the frequency domain and the compressed domain. In the frequency domain, an invertible neural network is designed to conceal secret message, where the modifications are directly written into the YUV file. In the compressed domain, data is hidden by modifying Coding Unit (CU) partition during the HEVC encoding process. The impacts of recompression and the available embedding positions are presented through detailed statistics and analysis. Experimental results demonstrate that the proposed algorithm maintains high visual quality and security. Crucially, it ensures reliable and accurate extraction of the secret information even after the video has undergone recompression.