In practical applications, stego videos inevitably undergo video encoding during transmission, where modern codecs adopt differentiated compression strategies for different frame types. However, most existing video hiding methods fail to adequately account for such codec-specific behaviors, leading to degraded performance under real-world video compression scenarios. To address the limited robustness of current video steganography approaches in realistic encoding environments, this paper proposes a Codec-Aware Video Hiding (CAVH) framework that approximates the characteristic distortions of the real video encoding pipeline by incorporating frame-type-specific priors. The proposed framework is built upon an invertible neural network (INN) backbone and incorporates a frame-type-aware compression simulator (FTACS), which approximates codec-induced distortions by distinguishing between I-, P-, and B-frame type information, thereby reducing the distribution gap between training-time distortion simulation and real encoding conditions. In addition, an inference-time residual scaling mechanism is introduced to adaptively control the embedding perturbation magnitude, enabling a flexible trade-off between cover video quality and secret recovery robustness without retraining. Furthermore, a decoder-side conditional refinement module is designed to perform nonlinear restoration of the recovered secret video, where motion residual priors constructed via optical flow are leveraged to enhance reconstruction quality under compression artifacts. Extensive experiments conducted on the UCF101 and REDS datasets demonstrate that, under both H.264/AVC and H.265/HEVC encoding conditions, the proposed method enhances secret video reconstruction quality while maintaining a competitive level of visual imperceptibility of the cover video. Moreover, CAVH exhibits strong generalization capability across different coding standards and data distributions.
Lu et al. (2026) studied this question.