Introduction
Cloud-based video analytics offer significant computational advantages for tasks like surveillance and industrial monitoring. However, transmitting raw video data to third-party servers introduces severe privacy risks. Traditional solutions often involve a trade-off: anonymization techniques like blurring or downsampling protect privacy but degrade AI recognition accuracy, while standard encryption prevents AI processing altogether without decryption. This paper introduces a novel framework that achieves privacy-preserving action recognition with zero loss in accuracy.
Methodology
The proposed approach consists of two integrated components: Learnable Cube-based Video Encryption (LCVE) and ViT Scrambling.
Learnable Cube-based Video Encryption (LCVE)
Unlike frame-based encryption, LCVE treats video as a 4D spatio-temporal volume. The process involves two primary stages:
- Cube Shuffling: The video is partitioned into spatio-temporal cubes (e.g., $2 \times 16 \times 16$ blocks), which are then randomly shuffled throughout the video volume.
- Pixel Shuffling: Within each individual cube, RGB pixel values are shuffled based on a secret key.
This multi-level shuffling results in a video that appears as visual noise to human observers and unauthorized AI models, while significantly expanding the key space to approximately $3.5 \times 10^{8560}$, rendering brute-force attacks computationally infeasible.
ViT Scrambling
The core innovation lies in the synergy between LCVE and Vision Transformers (ViTs). Because ViTs process images as sequences of patches, the researchers developed a method to "scramble" the model's internal weights to match the encrypted input. By rearranging positional embeddings and linear projection weights according to the encryption key, the model's first-layer output remains mathematically identical to that of a model processing unencrypted video. This eliminates the need for retraining on encrypted data.
Experimental Results
The framework was evaluated across seven diverse datasets, including Kinetics400, SSV2, and privacy-focused benchmarks like the KTH Dataset.
The results demonstrate that LCVE maintains 80.2% accuracy on the Kinetics400 dataset—identical to the performance on plain video. Simultaneously, privacy protection is significantly enhanced; actor identification accuracy on the KTH dataset dropped from 98.8% to just 4.0% when using LCVE. This outperforms traditional "Extreme Low-resolution" methods, which typically suffer from substantial accuracy drops (often 20-30%) while providing weaker privacy guarantees.
Conclusion
By exploiting the permutation invariance of Vision Transformers, the LCVE framework provides a robust solution for secure cloud-based video analytics. While currently optimized for classification tasks and Transformer architectures, the method sets a new benchmark for "zero-utility-loss" privacy-preserving AI. Future work may extend these principles to dense-prediction tasks like segmentation and other neural architectures.
Summary updated