LENS: A graph-based ciphertext semantic retrieval scheme with lightweight representations

With the rise of cloud outsourcing and pre-trained language model-driven semantic retrieval, high-quality semantic search over encrypted data remains challenging. Existing schemes either depend on keywords or coarse representations that cannot capture complex semantics, or use high-dimensional dense vectors with linear scanning, causing prohibitive computation. Although tree indexes reduce latency, their rigid partitioning may miss neighbors. Meanwhile, encrypted graph indexes still lack explicit leakage characterization and security adaptation. To address these issues, we propose a ciphertext semantic retrieval scheme that jointly integrates robust semantic representation learning, lightweight vector compression, and graph-based navigation. Specifically, we construct low-dimensional compact semantic vectors for encrypted computation through hard-negative-driven sample construction, global-local collaborative optimization, and multi-level self-distillation. We then build an encrypted graph index whose retrieval workflow is isomorphic to hierarchical navigable small world graph (HNSW) search. We further characterize the leakage boundary of graph-based retrieval and prove chosen-keyword attack security under an IND-CPA secure inner-product primitive. Experiments on the TREC corpus demonstrate a better trade-off among query efficiency, retrieval quality, and scalability than scan-based and tree-based baselines. These results suggest that high-performance ciphertext semantic retrieval requires the joint design of representation learning, index topology, and security modeling.