DPV-VPP: A Dual-Layer Video Privacy Protection Model Design Combining Differential Privacy and Variational Autoencoder-Based Face Replacement

Abstract

To tackle the risk of visual content privacy leaks during video calls, the study proposes a two-layer protection method combining differential privacy with variational autoencoder-based face replacement. The first layer uses a 3D convolutional structure based on optical flow to extract temporal features. It also applies a block-level cropping perturbation to sensitive areas, ensuring frame consistency and effective privacy masking. In the second layer, a variational autoencoder is uses to replace faces, achieving natural transitions via semantic generation and boundary fusion. Experiments on the Celeb-DF dataset show the method achieves a 96.9% privacy protection success rate, 3.7% false negative rate, and 96.8% misdirection success rate against attacks. In simulated platform attack tests, the protection success rates against cross-site scripting injection and forged request attacks were 99.2% and 98.9%, respectively. In 95.1% of the test video frames, the system processing rate reached 30 frames per second, with a minimum CPU usage of 0.9% during processing. The results indicate that the method ensures visual privacy security while maintaining good real-time performance and deployment adaptability.