Optimizing encrypted search in the cloud using autoencoder-based query approximation

Abstract

Searching over encrypted data is critical for secure cloud services. However, encryption reduces search efficiency due to inability to directly compute on ciphertexts. This leads to an inherent tradeoff between privacy and usability that has motivated extensive research on enabling effective encrypted search. Recent work has explored using machine learning models like autoencoders to optimize similarity search under encryption. Autoencoders compress data into low-dimensional vectors capturing semantic information. Documents and queries can be encoded for efficient similarity computation without decryption. However, existing analysis remains limited in scope and scale. The research address these gaps through large-scale experiments and novel optimizations. Our work provides a rigorous evaluation of autoencoder-based query approximation for encrypted cloud search using real-world datasets. The research implement a general framework agnostic to model type, data modality, and encryption scheme. Comprehensive experiments are conducted on public datasets using cloud infrastructure to quantify accuracy, efficiency, and scalability. The research present an extensive study on optimizing encrypted search through query approximation with autoencoders. Our research contributes a systematic analysis of how different architectural and training choices impact performance. Additional novel techniques proposed include quantization to reduce computation and homomorphic encryption to prevent information leakage. Our work is comprehensively benchmarked against alternative methods to quantify gains. The open-source implementation enables further research into optimized neural encrypted search.

Received 15/11/2023

Revised 23/03/2024

Accepted 25/03/2024

Published Online First 20/06/2024