Federated Vector Similarity Search

Introduction

The rapid growth of large language models (LLMs), multimodal AI, and deep learning has made high-dimensional dense vectors a fundamental representation for modern information retrieval and knowledge management. As vector data becomes increasingly distributed across organizations, data federation provides a practical approach to enabling cross-domain retrieval without directly centralizing raw data. This is particularly important in privacy-sensitive domains such as healthcare and finance, where regulations including GDPR, CCPA, and PIPL impose strict constraints on data sharing and storage.

However, federated high-dimensional vector search introduces substantial computational and communication overhead. The curse of dimensionality and distance concentration reduce the efficiency of conventional indexing structures, while approximate nearest neighbor search (ANNS) methods such as IVF and HNSW require carefully configured search parameters to balance recall and efficiency. In existing federated retrieval systems, fixed search parameters and uniform candidate quotas are often assigned to all nodes, regardless of their data relevance, index type, data distribution, or availability. Such strategies lead to redundant local searches and unnecessary candidate transmission, especially under non-IID data distributions, limited bandwidth, and client churn.

To address these limitations, we propose federated optimized search (FedOS), a dynamic optimization framework for federated ANNS. For each query, FedOS employs a lightweight two-stage prediction pipeline to estimate the required candidate quota for each node and determine node-specific search parameters according to its local indexing method, such as IVF or HNSW. By jointly optimizing candidate allocation and local search configurations, FedOS reduces unnecessary index traversal and network transmission, thereby improving retrieval latency and overall system throughput.

We further extend FedOS into secure-FedOS (S-FedOS) by integrating trusted execution environments, differential privacy, and oblivious sorting to mitigate side-channel leakage and network traffic volume analysis. Extensive experiments on multiple large-scale, high-dimensional vector datasets show that FedOS and S-FedOS consistently maintain high recall while substantially improving throughput under identical privacy constraints. The proposed frameworks also remain effective under challenging non-IID data distributions and unstable client availability, outperforming state-of-the-art federated retrieval baselines.