Robust Deep Video Watermarking Against Camera Recording Attacks via Temporal Distortion Modeling and Synchronization

Camera recording attacks pose a serious challenge to video watermarking due to the complex spatiotemporal distortions introduced during the playback–recording process. In particular, frame-rate mismatch causes temporal resampling and severe desynchronization, making embedded watermarks difficult to reliably decode. Existing deep video watermarking frameworks predominantly emphasize robustness against spatial or digital distortions but do not explicitly model the temporal distortions introduced by camera recording, resulting in substantial degradation in watermark extraction. To address this issue, we propose RecMark, a deep video watermarking framework that integrates differentiable spatiotemporal distortion modeling with learned frame synchronization. Specifically, we introduce a differentiable distortion layer that simulates spatial degradations and temporal desynchronization caused by mismatched playback and recording frame rates. A dedicated synchronization watermark is embedded in the frequency domain to provide frame-level synchronization cues and reconstruct temporally aligned frame groups, while the copyright watermark is embedded in the spatiotemporal domain through an encoder equipped with a Spatiotemporal Fusion Module (SFM). Extensive experiments demonstrate that RecMark achieves high extraction accuracy under camera recording attacks while preserving high perceptual video quality.