Privacy-preserving Image Retrieval Based on Additive Secret Sharing in Cloud Environment

Abstract

Abstract With the growth of cloud computing, an increasing number of resource-constrained data owners tend to store their images on the cloud, and cloud servers provide image retrieval services for users. However, this poses significant privacy leakage risks. Privacy-preserving image retrieval can prevent privacy leakage during retrieval, but existing schemes have the problem of low search accuracy. Additionally, existing solutions usually utilize a single cloud server to provide retrieval services, which suffer from problems such as a single point of failure and a greater risk of privacy leakage. This is because the cloud server is honest but curious, and it may peek at users' data out of curiosity or interest.With the proliferation of cloud service providers, image retrieval services can be supported collaboratively by multiple cloud servers. However, complex encryption techniques and frequent interactions between servers significantly degrade the performance of the overall search process.This paper proposes a lightweight three-party computation (3-PC) image retrieval strategy based on additive secret sharing in a cloud environment.Specifically, the data owner extracts the features of images using a pre-trained CNN model and constructs an index for the features using the K-means hierarchical index tree. After splitting and outsourcing the images, CNN model, and index tree to three cloud servers, the data owner remains offline. The three cloud servers process the user's query request and utilize a secure computation protocol based on secret sharing. This protocol ensures that the parameters of the neural network and the extracted query image features are hidden from the cloud servers and other attackers. Additionally, a secure feature similarity computation and distance ranking protocol is designed to avoid privacy leakage during the search process.Security analysis and experimental results demonstrate the security and retrieval performance of the scheme.