An Efficient Protocol for Private Face Recognition Using Post-Quantum Secure Encryption

We propose an efficient private face recognition protocol using CKKS fully homomorphic encryption that addresses SIMD slot waste in per-embedding CipherFace baselines. By row-packing multiple face embeddings into contiguous blocks within a single ciphertext and applying a mask-free block-isolated hierarchical reduction, our method enables parallel distance computation over packed embeddings. Experiments on the LFW dataset with FaceNet, FaceNet512, and VGG-Face embeddings demonstrate significant speedups over the CipherFace baseline, with Hamming distance on binarized embeddings offering additional efficiency while maintaining comparable accuracy at d=512.