CASTNet: a content-adaptive spatio-temporal network for robust video hiding

Abstract

Existing robust video hiding methods commonly apply a sample-independent distortion distribution and model temporal dependencies at a fixed range, leaving compression supervision weakly coupled to the texture and motion characteristics that affect practical video coding. We propose CASTNet, whose central idea is to couple content-conditioned distortion supervision with invertibility-compatible multi-range temporal representation, rather than merely appending independent modules to an invertible backbone. The Content-Aware Adaptive Compression Simulator (CA-ACS) maps texture-complexity and motion-intensity descriptors to sample-specific mixture weights over differentiable distortion branches, so that videos with different content characteristics receive different robustness perturbations during training. The Multi-Scale Spatio-Temporal Fusion (MSF) mechanism is embedded inside the transformation subnetworks of affine coupling blocks and is recomputed from variables available in both forward hiding and inverse recovery. It therefore provides complementary temporal-range features for dense video hiding without introducing external side information that would invalidate the inverse process. Established attention operators are used to selectively refine these conditional features, while progressive distortion injection separates the learning of a stable hiding–recovery mapping from subsequent compression adaptation. Under the evaluated UCF-101 training protocol and codec configurations, the experimental results provide evidence that this coordinated design improves stego-video quality and secret-video recovery over representative image/video hiding baselines and the recent invertible video-hiding method LF-VSN under no compression and representative H.264/H.265 conditions. Additional tests cover VP9, cascaded re-encoding, cross-codec transcoding, and a controlled UCF-101-to-REDS zero-shot transfer in which clip length, spatial resolution, payload, and codec settings are held fixed. Computational profiling further quantifies the additional cost of MSF/AEM and confirms that CA-ACS is a training-only component with no inference-time module or parameter overhead. A dedicated cover–stego classification evaluation with spatial and spatio-temporal residual detectors shows lower detectability for CASTNet than VHNet at the native payload before and after H.264 coding, although detection remains above chance and no claim of statistical or cryptographic security is made.