Music Retrieval via Additive Homomorphic Encryption

A new research paper from arXiv (2508.07044) proposes using additive homomorphic encryption for music information retrieval to balance privacy and efficiency. The authors note that modern music retrieval relies on vector embeddings, which can be copied or used to train generative models when shared. Fully homomorphic encryption is impractical due to ciphertext multiplication and bootstrapping. They observe that when only one operand is encrypted, similarity search reduces to ciphertext addition and ciphertext-plaintext multiplication, supported by Paillier or CKKS restricted to additive operations. The study implements two music-specific inference attacks, quantifies privacy-utility tradeoffs, introduces structure-aware additive primitives with learned per-block weighting, and demonstrates across four audio datasets that additive search preserves utility while reducing computational overhead.