<p>Deep learning has significantly advanced digital watermarking, providing greater robustness, adaptability, and imperceptibility compared to traditional techniques. While image watermarking using deep networks is relatively mature, video watermarking remains a more complex and underexplored domain. This paper introduces a novel video watermarking technique based on Convolutional Neural Networks (CNNs), specifically a ResNet architecture enhanced with channel and spatial attention (CBAM), to achieve robust and imperceptible watermark embedding and extraction. The proposed method operates on mosaic images generated from video sequences, enabling spatially coherent temporal insertion. Watermark information is embedded into geometrically aligned regions representing temporally consistent scene content. The full pipeline comprises: a preprocessing module for mosaic generation, an attention-augmented ResNet-based embedder with an adaptive strength factor for content-aware watermark modulation, a simulated attack layer during training to improve resilience, and a multi-scale decoder network for blind watermark retrieval. Data augmentation is further applied to enhance generalization across unseen distortions. Extensive experiments demonstrate that the proposed method provides high visual quality, reaching up to 45 dB PSNR, while maintaining a BER = 0.000 in the absence of attacks. Under various distortions, including noise, geometric transformations, compression, and collusion, the method achieves consistently low BER values. Compared to baseline image and video watermarking techniques, our model offers a superior trade-off between invisibility, robustness, and capacity, underscoring its potential for practical and secure video content protection.</p>