Privacy-Preserving Retrieval Scheme Over Medical Images Based on Vision Transformer

Online pathological diagnosis is essential for improving disease detection and advancing telemedicine but faces challenges related to data privacy and security. Therefore, we propose SecViT, a medical image retrieval scheme that uses secret-sharing technology to deploy a Vision Transformer model across two separate cloud servers via interactive protocols for secure image storage and processing. Significantly, we enhance processing efficiency by improving the comparison protocol. Additionally, to achieve secure and efficient division computations, we design secure scaling and iterative protocols based on the coordinate rotation digital computer algorithm. Utilizing these improved protocols, the advantages of CNNs and Transformers can be effectively combined to achieve efficient and privacy-preserving image retrieval. Finally, empirical experiments prove that it outperforms state-of-the-art privacy-preserving solutions for medical images regarding computational cost, communication overhead, and retrieval accuracy.