CFE-PPAR: Compression-Friendly Encryption for Privacy-Preserving Action Recognition
Researchers have introduced CFE-PPAR, a novel compression-friendly encryption method designed for privacy-preserving action recognition (PPAR) in video analysis. While existing encryption-based PPAR strategies offer strong privacy protection, they typically suffer from significant drops in recognition performance and visual quality when videos are compressed using standard codecs. To overcome this limitation, the proposed CFE-PPAR framework allows video transformers to directly recognize activities from encrypted videos by utilizing model parameters transformed with the same secret keys used for encryption. This approach ensures that privacy is maintained without compromising analytical accuracy under compression. Experimental results demonstrate that CFE-PPAR outperforms previous methods on the UCF101 and HMDB51 datasets when subjected to Motion-JPEG and H.264 compression standards. This development represents a significant advancement in computer vision, enabling secure and efficient human activity understanding in scenarios where bandwidth constraints necessitate video compression, such as cloud-based surveillance or remote monitoring systems.
Wire timeline
CFE-PPAR: Compression-Friendly Encryption for Privacy-Preserving Action Recognition
Researchers have introduced CFE-PPAR, a novel compression-friendly encryption method designed for privacy-preserving action recognition (PPAR) in video analysis. While existing encryption-based PPAR strategies offer strong privacy protection, they typically suffer from significant drops in recognition performance and visual quality when videos are compressed using standard codecs. To overcome this limitation, the proposed CFE-PPAR framework allows video transformers to directly recognize activities from encrypted videos by utilizing model parameters transformed with the same secret keys used for encryption. This approach ensures that privacy is maintained without compromising analytical accuracy under compression. Experimental results demonstrate that CFE-PPAR outperforms previous methods on the UCF101 and HMDB51 datasets when subjected to Motion-JPEG and H.264 compression standards. This development represents a significant advancement in computer vision, enabling secure and efficient human activity understanding in scenarios where bandwidth constraints necessitate video compression, such as cloud-based surveillance or remote monitoring systems.
cs.AI updates on arXiv.org