<p>Steganography secures secret information by hiding it in plain sight within common digital media. Due to the increasing multimedia data, digital videos provide abundant source of high-capacity cover media, creating a strong practical demand for utilizing video streams to securely conceal and transmit secret images. Most existing studies focus on unimodal methods and lack exploration of cross-domain concealment between videos and images. Unlike unimodal approaches, cross-domain steganography faces two challenges: bridging the spatio-temporal mismatch between images and videos, and enhancing the robustness of invertible mappings against lossy transmission. This paper proposes a novel cross-domain steganography framework for hiding images within videos, namely VSHI-Net. Specifically, to address the spatio-temporal mismatch, a content-aware Memory Bank module is constructed. This module extracts and stores the deep feature representations of the cover video frames, enabling the adaptive retrieval of the most compatible set of video frames to serve as carriers. Based on this, a sliding window cropping strategy is further introduced to segment the secret image into patches. These patches are then precisely aligned with the optimal set of cover frames selected by the Memory Bank module, reducing the spatio-temporal mismatch. Furthermore, to reduce the vulnerability of invertible mappings, a Noise-Based Invertible Neural Networks backbone is designed. By introducing noise as an auxiliary input, the model aims to improve resistance against perturbations during lossy transmission. Extensive experiments demonstrated that, compared to other state-of-the-art methods, our VSHI-Net achieves competitive performance in terms of invisibility, recovery accuracy, and security.</p>