End-to-End Secure Image Retrieval via Self-Supervised Learning in Cloud Computing
Zhixun Lu, Yan‐Feng Chen, Qihua Feng, Jing Liang, Peiya Li · 2023
Joint image encryption and retrieval technology is a promising research direction, which aims to ensure the data security and searchability in cloud environment. However, inappropriate image cryptosystems and inefficient feature extractors will lead to poor performance since the images’ features are distorted and hard to analyze. In this paper, we propose a novel privacy-preserving image retrieval scheme which is able to overcome the above two defects. To be specific, image content is encrypted by block rotation, new orthogonal transforms and block permutation during JPEG compression, and we design an end-to-end deep learning based self-supervised secure image retrieval model named Simplified Restormer Network (SRN) as the feature extractor, which can effectively extract cipher-images’ local-global representations. After encryption, cipher-images can be directly input into our SRN model without any processing. In addition, considering the existence of databases without labels, we propose two model training schemes to train our SRN model, supervised and self-supervised, respectively. The experimental results show that our retrieval method can achieve superior performance than other state-of-the-art retrieval methods, and our encryption algorithm can ensure cipher-image compression-friendly as well as no information leakage.