Deep adaptive cryptographic video watermarking using hybrid DCT–SVD and CNN-based adaptive embedding

As the use of digital multimedia distribution, which includes cloud service providers, social networks and streaming services, has grown rapidly, there has been increasing demand for robustly protecting copyright, verifying content authenticity and determining ownership. Previous watermarking methods may be difficult to balance the watermarking goals—imperceptibility, robustness, and security, and traditional cryptographic approaches can provide security during the transmission but they cannot support the ownership information after de-cryption. To address these issues, the Deep Adaptive Cryptographic Video Watermarking Framework is proposed in this paper, which combines the transform domain based watermark embedding technique with adaptive embedding approach based on deep learning and chaotic encryption. The framework makes use of the Discrete Cosine Transform (DCT) together with Singular Value Decomposition (SVD) for embedding the encrypted watermark information into the mid frequency coefficient, the strength of embedding information is predicted adaptively based on the local spatial-frequency characteristics, and a light-convolutional network is used for the prediction. A noisy channel model for the residual neural decoder recovers the embedded watermark from corrupted video frames to be more resilient for compression, Gaussian noise, cropping, scaling and frame manipulation events. Choa’s chaotic key-generation mechanism is used in a secure hash-assisted way, which is used independently to generate keys for preventing the tampering of image watermark using the neural network. By simulation testing of the system under various attacking conditions, (both signal processing and geometric attacks), the system is evaluated to show that the proposed framework meets high visual qualities and the watermark can be recovered under different attacks and various watermark payloads. The results also show the well-known trade-off between watermark capacity and its imperceptibility and extraction accuracy, and a higher payload results in a moderate decrease of the PSNR and consequently, an increase of the BER in the case of severe attacks. Comparative analysis demonstrates that the proposed framework offers a well-balanced performance among the visual quality, robustness, and payload security as compared with the conventional transform-domain and existing watermarking approaches using deep learning. All of the above makes the proposed scheme can be used as a means of secure authentication for multimedia, copyright protection and protected video distribution.