pbins: private bins for top-k semantic search over encrypted data using transformers

Traditional searchable encryption constructions are often based on Term-Frequency and Inverse Document Frequency (TF-IDF) to retrieve relevant documents. These vectors are sensitive to the document length, sparse and lacks semantic understanding. Additionally, to address the growing privacy challenges, we propose pbins a novel design for achieving privacy preserving semantic-aware multi-keyword ranked search over encrypted data using “Private bins”. pbins is a method-agnostic encapsulation designed for indexes based on vector embeddings, enabling document retrieval without compromising privacy. Our approach leverages SBERT (Sentence-BERT) like Transformers to capture semantically meaningful sentence embeddings. The private bins group semantically similar documents together based on relevance scores thus allowing users to retrieve information without compromising data confidentiality. Private Bins can be updated dynamically without requiring a complete index reconstruction or uploading new structures to the cloud. Furthermore, pbins supports range-encoded and ranked search allowing users to retrieve the top-k documents efficiently while preserving privacy. The robustness of our approach is substantiated through a thorough security analysis that demonstrates correctness and forward privacy. Experiments on real-world news group dataset demonstrates the performance of pbins in terms of accuracy and efficiency.