Private Information Retrieval (PIR) permits clients to query data entries from a public database hosted on untrusted servers while preserving client privacy. Traditional PIR models suffer from high computation and/or bandwidth overhead due to linear database processing. Recently, Online-Offline PIR (OO-PIR) has been proposed to improve PIR practicality by precomputing query-independent materials to accelerate online access. While state-of-the-art OO-PIR schemes (e.g., S&P’24, CRYPTO’23) successfully reduce online processing cost to sublinear levels, they still impose substantial bandwidth and storage burdens on the client, especially when operating on large databases. In this paper, we propose Pirex, a new two-server OO-PIR scheme with semi-honest security that offers minimal client-side inbound bandwidth and storage costs while retaining sublinear processing efficiency. The Pirex design is simple, with most operations being naturally low-cost and streamlined (e.g., XOR, PRF, modular arithmetic). We have fully implemented Pirex and evaluated its real-world performance using commodity hardware. Our results show that Pirex outperforms existing OO-PIR schemes by at least two orders of magnitude. With a 1 TB database, Pirex takes only 55 ms to retrieve a 4 KB entry, compared to 9–30 seconds for state-of-the-art approaches. For practical databases with billions of 4 KB entries, Pirex requires just 16 KB of inbound bandwidth—up to three orders of magnitude more efficient.

Federated learning (FL) allows multiple distrustful clients to collaboratively train a machine learning model. In FL, data never leaves client devices; instead, clients only share locally computed gradients with a central server. As individual gradients may leak information about a given client’s dataset, secure aggregation was proposed. With secure aggregation, the server only receives the aggregate gradient update from the set of all sampled clients without being able to access any individual gradient. One challenge in FL is the systems-level heterogeneity that is quite often present among client devices. Specifically, clients in the FL protocol may have varying levels of compute power, on-device memory, and communication bandwidth. These limitations are addressed by model-heterogeneous FL schemes, where clients are able to train on subsets of the global model. Despite the benefits of model-heterogeneous schemes in addressing systems-level challenges, the implications of these schemes on client privacy have not been thoroughly investigated.