Paper 2025/1479

PicoGRAM: Practical Garbled RAM from Decisional Diffie-Hellman

Abstract

Making 2-party computation scale up to big datasets is a long-cherished dream of our community. More than a decade ago, a line of work has implemented and optimized interactive RAM-model 2-party computation (2PC), achieving somewhat reasonable concrete performance on large datasets, but unfortunately suffering from $\widetilde{O}(T)$ roundtrips for a $T$-time computation. Garbled RAM promises to compress the number of roundtrips to $2$, and encouragingly, a line of recent work has designed concretely efficient Garbled RAM schemes whose asymptotic communication and computation costs almost match the best known interactive RAM-model 2PC, but still leaves $({\sf poly})\log\log$ gaps. We present ${\sf PicoGRAM}$, a practical garbled RAM (GRAM) scheme that not only asymptotically matches the prior best RAM-model 2PC, but also achieves an order of magnitude concrete improvement in online time relative to interactive RAM-model 2PC, on a dataset of size $8$GB. Moreover, our work also gives the first Garbled RAM whose total cost (including bandwidth and computation) achieves an optimal dependency on the database size (up to an arbitrarily small super-constant factor). Our work shows that for high-value real-life applications such as Signal, blockchains, and Meta that require oblivious accesses to large datasets, Garbled RAM is a promising direction towards eventually removing the trusted hardware assumption that exist in production implementations today. Our open source code is available at https://github.com/picogramimpl/picogram.