Preprint reports faster encrypted text retrieval through batched computation

Researchers report reducing encrypted text retrieval latency from 222.028 to 5.168 seconds for a single query over 7,128 indexed items. Their arXiv preprint examines health-misinformation classification and semantic retrieval while protecting potentially sensitive user text. The optimized implementation uses CKKS homomorphic encryption, which supports computation on encrypted data, and packs multiple similarity calculations together. Communication overhead fell from 6.11 GB to 2.13 MB. A separate hybrid classifier achieved accuracy of 0.971 and macro-F1 of 0.950, with identical reported metrics for plaintext and encrypted inference. The retrieval speed comparison is against the researchers' naive encrypted implementation.