Homomorphic Feature Classification for Confidential Nearest-Neighbor Search

The system segments the classification task into two parts: a complex deep neural network embedding module that runs locally on the user's device to extract feature vectors, and a simpler classification module that runs on the server using homomorphic encryption. This segmentation allows the computationally intensive part to be performed locally without encryption overhead, while the server only handles lighter encrypted operations.

The embedding model is trained and fixed in advance during an offline phase. The complex neural network weights are predetermined and stored, so during online classification, only the simpler nearest-neighbor search needs to be performed on encrypted data, significantly reducing the computational complexity of encrypted operations.

Homomorphic encryption acts as an intermediary mechanism that enables confidential computation without degrading performance. Instead of adding noise to reference data, the system uses cryptographic primitives to protect the classification results while maintaining the integrity and accuracy of the original data for recognition purposes.

The patent replaces the mechanical approach of noise-making (which directly interferes with data quality) with a cryptographic approach (homomorphic encryption) that protects confidentiality through mathematical transformations without affecting the underlying data quality or recognition performance.

The recognition system is divided into a fixed embedding module (trained once offline) and a flexible classification module (can be updated online). When new faces need to be recognized, only the simple classification module needs to be updated with new reference vectors, rather than retraining the entire complex neural network, thus improving system maintainability.

The complex feature extraction capabilities are prepared in advance through offline training of the embedding model. This preliminary action creates a robust foundation that can accommodate new classification tasks by simply adding new reference vectors to the classification module, avoiding the need for complete retraining.

The system segments operations into local encrypted feature extraction (using pre-trained embedding models) and server-side encrypted classification. The embedding models are optimized to work efficiently with encrypted inputs, and the server performs only simple distance calculations in the encrypted domain, maintaining good classification efficiency.

The system changes the parameter space by projecting high-dimensional encrypted input data into a lower-dimensional embedding space using pre-trained models. This dimensionality reduction maintains the essential features needed for classification while reducing the computational burden of subsequent encrypted operations, thus improving efficiency.