A H.265/HEVC Fine-Grained ROI Video Encryption Algorithm Based on Coding Unit and Prompt Segmentation
A new method uses AI prompt segmentation to encrypt only sensitive parts of a video, like faces or documents.
A research team led by Xiang Zhang has developed a novel algorithm for selectively encrypting sensitive parts of videos, a critical need for fields like medicine and military surveillance. The core innovation is moving away from large, imprecise 'Tile' units to a much finer-grained approach. By combining AI-powered prompt segmentation with the H.265/HEVC video codec's internal structure, the algorithm can pinpoint and map Regions of Interest (ROI)—such as a person's face or a confidential document—down to the smallest 8x8 pixel block, known as a Coding Unit (CU).
This precision is paired with a selective encryption scheme that distorts specific syntax elements within these targeted CUs, effectively scrambling the visual data. A major technical hurdle overcome is 'encryption diffusion,' where scrambling one block corrupts the prediction for neighboring blocks, causing visual artifacts. The team solved this by applying Pulse Code Modulation (PCM) mode and motion vector restrictions to isolate the encrypted CUs. The result is a method that can accurately encrypt only the sensitive portions of a video frame without degrading the rest of the image or causing blurring or blocky artifacts, enabling secure sharing of videos where context must remain visible but specific details must be hidden.
- Uses AI prompt segmentation to identify sensitive video regions (ROI) for encryption.
- Encrypts at a precise 8x8 Coding Unit level, a major leap from older, blockier 'Tile' methods.
- Solves encryption 'diffusion' artifacts using PCM mode and MV restrictions to isolate scrambled blocks.
Why It Matters
Enables secure video sharing in healthcare and defense by hiding only sensitive details, keeping the rest of the scene usable.