Homomorphic Encryption for Privacy-Preserving Feature Store Queries in Federated Learning

Abstract

This paper proposes partly homomorphic Paillier encryption for cross-administrative  federated learning feature store privacy and searches. As more companies employ safe and  legal data, we provide a cryptographic approach to linearly aggregate attributes without  revealing plaintext values. The work provides a secure computing pipeline for in situ  encrypted feature vector searches and aggregation. Ciphertext packing reduces computation  and storage. Distributed key management ensures security and scalability. Performance must  be traded to scale data encryption across GPU-accelerated nodes and ARM-based edge  devices. Hierarchical federated learning topologies with millions of encrypted data  consolidate with acceptable latency and accuracy. The proposed system has high throughput,  accuracy, and collaborative machine learning privacy.