STAR-Net: spatio-temporal invertible network for video watermarking against composite attacks

Video watermarking is crucial for digital media copyright protection. Existing invertible neural network (INN)-based methods predominantly rely on passive zero-padding to compensate for missing latent variables during extraction, which often leads to significant performance degradation under real-world composite attacks due to the irreversible loss of high-frequency information. To address this, we propose STAR-Net, a framework that shifts the paradigm from passive filling to active spatio-temporal prior reconstruction. Our approach introduces a Texture-Consistent Semantic Embedding (TCSE) module to project discrete watermark bits into a spatial representation aligned with the host video’s high-frequency wavelet subbands, thereby enhancing embedding invisibility and mitigating semantic mismatch. Furthermore, to overcome the bottleneck of blind extraction, we design a Spatio-Temporal Selective-Kernel Latent Predictor (ST-SKLP) equipped with a multi-scale mechanism to actively reconstruct missing high-frequency residual priors from degraded videos. To bolster robustness, we construct a composite attack layer integrating a Gradient-Aware Perturbation Engine (GAPE) and a Differentiable High-Frequency Quantization Surrogate (DHQS). The GAPE generates adversarial noise tailored to maximize extraction error, simulating severe channel distortions. Extensive experiments on UCF-101 and Kinetics-600 demonstrate that STAR-Net achieves superior robustness and visual quality, effectively reconciling the trade-off between imperceptibility and extraction accuracy even under extreme compression.