Image Feature Extraction in Encrypted Domain With Privacy-Preserving Hahn Moments

With outsourcing huge amount of images to cloud, it can reduce the local storage and computational overhead of resource-limited users, but the security concerns still impede the adoption of outsourced image processing services. Fortunately, image processing in the encrypted domain can handle this problem very well and has already attracted considerable attention due to its ability to protect valuable image information from being leaked to untrusted parties. However, the existing schemes incur high computational complexity, and none of them can be used as both global feature and local feature. To this end, we present a method to implement Hahn moments in the encrypted domain by using Somewhat Homomorphic Encryption (SHE) in this paper, named Privacy-Preserving Hahn Moments (PPHM). First, a mathematical framework is proposed to implement the PPHM and image reconstruction in the encrypted domain. Then, the detailed theoretical analysis about data expansion and quantization errors shows that plaintext Hahn moments and plaintext image reconstruction can be implemented by utilizing PPHM over encrypted image. Moreover, security analysis shows that the PPHM can guarantee the image content security. Theoretical analysis and experimental results demonstrate that the PPHM scheme can greatly reduce the computational complexity compared with the discrete cosine transform and the discrete wavelet transform in the encrypted domain. In addition, experimental results show that the PPHM has a better performance in both image reconstruction and image recognition. Furthermore, the suitable value of a scaling factor, which can provide guidance to users, is specified for image reconstruction and image recognition, respectively.