Paper 2026/1543

Efficient Private Filtering and Aggregation for Weighted Set Intersection via Oblivious Encrypted Weight Transfer

Abstract

Private Set Intersection (PSI) enables parties to compute the intersection of their input item sets while preserving privacy. In many real-world applications, however, each item is accompanied by a sensitive weight, and the ability to privately compute over such weights is crucial. Existing research in this direction is fragmented and driven by application-specific goals, with representative examples including PI-Sum (computing the sum of weights over the intersection), inner-product Private Join and Compute (computing the inner product of weight vectors over the intersection), and Item with Maximum Weight Sum (identifying the intersection item with the maximum combined weight). In this work, we propose a unified framework for private computation on weighted set intersection. We formalize \textit{Private Filtering and Aggregation for Weighted Set Intersection} (PFA-WSI) as an ideal functionality parameterized by a joint scoring function $f$ and a predicate $P$, supporting two output modes: (i) \emph{predicate-filtered output}, which reveals a predicate-selected subset of intersection items, and (ii) \emph{aggregated output}, which reveals only aggregate statistics over matched items. By instantiating $f$ and $P$ appropriately, PFA-WSI captures deployed and studied tasks such as PI-Sum, inner-product PJC, and IMWS, and also accommodates richer metrics arising in practice, such as $L_1$- and $L_2$-type distance statistics on matched item weights. To realize PFA-WSI efficiently, we introduce a novel core building block, Oblivious Encrypted Weight Transfer (OEWT), which enables a receiver to obtain encryptions of the sender's weights for intersection items and random-looking ciphertexts otherwise. Building on OEWT and additively homomorphic encryption, we present modular protocol constructions for different instantiations of $f$ and for both output modes. We prove simulation-based security in the semi-honest model and provide detailed communication and computation analyses. Our experiments show that our constructions scale to million-sized sets with practical performance that matches or surpasses the state-of-the-art.

Note: This is the full version of the CCS'26 paper with the same title, including all theorem proofs.